Modèles d’IA

Claude Haiku 4.5

Anthropic

Recommandé

La référence rapport qualité-prix d'Anthropic à $1/$5 par 1M tokens — le palier rapide et économique auquel les concurrents se comparent, avec un contexte de 200K et une sortie de 64K. Idéal pour les charges à fort volume sensibles à la latence : classification, extraction, résumé, routage. Passez à Opus 4.8 ou Sonnet 5 lorsqu'un raisonnement profond en plusieurs étapes est nécessaire.

Entrée

$1 / 1M tokens

Sortie

$5 / 1M tokens

Contexte

200K ctx; legacy, Haiku 5.5 is current

Claude Opus 4.8

Anthropic

Recommandé

Claude Opus 4.8 is Anthropic's previous-generation Opus, now listed as legacy: 1M-token context, 128K output, $5/$25 per 1M tokens, retirement not before 28 May 2027. It remains our pick for review-heavy work such as code review and constrained refactoring, where deliberate pacing beats persistence. New workloads should start on Claude Opus 5.5 at $4/$20, which Anthropic now recommends.

Entrée

$5 / 1M tokens

Sortie

$25 / 1M tokens

Contexte

1M tokens

Claude Opus 5

Anthropic

Recommandé

Claude Opus 5 delivers near-Fable 5 intelligence at $5/$25 per 1M tokens, with a 1M-token context and 128K output. Anthropic now lists it as legacy: still available, retirement not before 24 July 2027, and Claude Opus 5.5 ($4/$20) is its recommended starting point. Recommended for teams already running Opus 5; new workloads should start on Opus 5.5, which is cheaper per token.

Entrée

$5.00 / 1M tokens

Sortie

$25.00 / 1M tokens

Contexte

1M ctx; 128K out; legacy; fast $10/$50

Claude Sonnet 5

Anthropic

Recommandé

Anthropic's most agentic Sonnet at launch delivers performance close to Opus 4.8 at 40 percent of the Opus rate. The $2/$10 per 1M token price is the standard rate: Anthropic cancelled the increase to $3/$15 that was scheduled for 1 September 2026. The newer tokenizer produces about 30 percent more tokens for the same text, so recount budgets before migrating from Sonnet 4.6. Anthropic now lists Sonnet 5 as legacy; Claude Sonnet 5.5, at the same $2/$10, is the current Sonnet, so start new work there.

Entrée

$2 / 1M

Sortie

$10 / 1M

Contexte

1M tokens; batch $1/$5; cache read $0.20

DeepSeek V4 Flash

DeepSeek

Recommandé

DeepSeek retired V4 Flash from its API on 10 September 2026: the deepseek-v4-flash id is still accepted, but its requests are served by V4.1-Flash and billed at the Flash rate of $0.15 input and $0.60 output per 1M tokens off-peak. The MIT-licensed weights (284B parameters, 13B active, 1M-token context) remain on Hugging Face, so this is now a frozen checkpoint for self-hosting and reproducible evals. For new API work, call deepseek-flash.

Entrée

$0.15 / 1M tokens

Sortie

$0.60 / 1M tokens

Contexte

Retired Sep 10; id serves V4.1-Flash. Peak 2x M-F

DeepSeek V4 Pro

DeepSeek

Recommandé

Strong price-performance on open MIT weights; the 2x peak window is 01:00-04:00 and 06:00-10:00 UTC on weekdays only, so weekends bill entirely off-peak.

Entrée

$0.66 / 1M tokens

Sortie

$1.98 / 1M tokens

Contexte

1M tokens; 2x price at weekday peak hours

Gemini 2.5 Pro

Google

Recommandé

Désormais le choix économique pour le contexte long : une baisse de prix de 40 % (juin 2026) a rendu les traitements complets de corpus accessibles aux budgets moyens. Il offre la plus grande fenêtre de contexte disponible, avec une intégration solide dans l'écosystème Google.

Entrée

$1.25 / 1M tokens

Sortie

$10 / 1M tokens

Contexte

1M tokens

Gemini 3.5 Flash

Google

Recommandé

La référence de Google en efficacité-coût pour l'automatisation agentique du poste de travail, à $1.50/$9 par 1M tokens avec un contexte d'environ 1M. Le Computer Use natif obtient 78,4 sur OSWorld-Verified — à 0,3 point de GPT-5.5 pour environ un tiers du coût. LA réserve à connaître : l'outil Computer Use est en préversion publique, pas encore dans l'API de production ; vérifiez la disponibilité de l'endpoint avant de l'intégrer.

Entrée

$1.50 / 1M tokens

Sortie

$9 / 1M tokens

Contexte

1M tokens

Gemini 3.5 Flash-Lite

Google

Recommandé

Gemini 3.5 Flash-Lite est le modèle le plus rapide et le moins cher de la série 3.5 de Google, à 350 tok/s et 0,30 $/2,50 $. Le saut générationnel par rapport à 3.1 Flash-Lite est le véritable sujet — il surpasse nettement son prédécesseur sur les principaux benchmarks agentiques et bat le modèle Gemini 3 Flash pleine taille sur les tâches de codage et de computer-use. Pour les développeurs qui font tourner à grande échelle des sous-agents agentiques, des pipelines de recherche et du traitement de documents, le rapport prix-performance de Flash-Lite est inégalé dans la gamme Google.

Entrée

$0.30 / 1M

Sortie

$2.50 / 1M

Contexte

1M context, 64K output, 350 tok/s

Gemini 3.6 Flash

Google

Recommandé

Gemini 3.6 Flash is Google's new default Flash-tier workhorse: better than 3.5 Flash on every benchmark and cheaper, at $0.75/$3.75 per 1M tokens through 2026-12-31, rising to $1.50/$7.50 from 2027-01-01. The 17% token-efficiency gain lowers per-task cost, and built-in Computer Use via the API makes it a credible agentic coding option. For Google-ecosystem developers it is the clear upgrade; budget for the January rate step.

Entrée

$0.75 / 1M (through 2026-12-31)

Sortie

$3.75 / 1M (through 2026-12-31)

Contexte

1M context, 64K output; $1.50/$7.50 from 2027-01-01

Gemini 3.7 Flash

Google

Recommandé

Gemini 3.7 Flash est le cheval de trait Flash le plus puissant de Google à ce jour pour le codage et les agents — une véritable montée en gamme par rapport à 3.6 Flash, à la moitié du prix précédent jusqu'en 2026. Il reste toutefois derrière la frontière sur les tâches terminal et d'utilisation d'ordinateur les plus difficiles. Une mise à niveau évidente pour les créateurs d'agents sensibles aux coûts et ancrés dans l'écosystème Google ; pas un remplaçant des modèles frontières pour le code.

Entrée

$0.75 / 1M

Sortie

$3.75 / 1M

Contexte

1M context, 64K output

Gemini 3.8 Flash

Google

Recommandé

Gemini 3.8 Flash is Google's strongest Flash workhorse yet — a genuine capability upgrade over the recommended 3.7 Flash at the same introductory price, with real gains across software engineering, agents and professional domains. The deciding caveat is Google's own: 3.8 "works harder," so per-task token burn runs well above 3.7, which remains the efficiency-first pick. Adopt 3.8 for long-horizon coding, agents and finance or legal work — benchmark your own workload before the intro price expires.

Entrée

$0.75 / 1M

Sortie

$3.75 / 1M

Contexte

1M context, 64K output

Gemini 3.8 Live Extended Thinking

Google

Recommandé

Gemini 3.8 Live Extended Thinking is the voice-agent model to start with: it tops Artificial Analysis' Speech to Speech Index and reasons and calls tools in the background while it keeps talking, at the same rate as the default Gemini 3.8 Live. The caveat is the Live API surface: no structured outputs, caching, code execution or batch. Right for teams building real-time voice agents that must solve multi-step problems mid-call.

Entrée

$0.75 / 1M (text)

Sortie

$4.50 / 1M (text)

Contexte

131K in, 65K out; audio $3/$12 per 1M

GPT-4.1

OpenAI

Recommandé

La cible de migration par défaut pour les workloads GPT-4 Turbo : même format de requête, coût inférieur, contexte de 1M tokens — convertissez les payloads function_call hérités au format tools et passez à la suite.

Entrée

$2 / 1M tokens

Sortie

$8 / 1M tokens

Contexte

1M tokens

GPT-5.5

OpenAI

Recommandé

GPT-5.5 is OpenAI's April 2026 general-purpose model at $5 input and $30 output per 1M tokens, with a 1M+ token context (922K in, 128K out). It is no longer OpenAI's only deployable frontier option: the GPT-5.6 family is on the public price list, and GPT-6 Sol lists at $2/$10. It stays recommended as a stable, widely supported production baseline, but new workloads should test GPT-6 Sol or GPT-5.6 Terra first.

Entrée

$5 / 1M tokens

Sortie

$30 / 1M tokens

Contexte

1M+ tokens (922K in / 128K out)

GPT-5.6 Sol

OpenAI

Recommandé

GPT-5.6 Sol is no longer OpenAI's best-value frontier model: its successor GPT-6 Sol, released 22 September 2026, lists at $2/$10 per 1M tokens against Sol's $4/$20. It stays recommended because its coding lead has third-party backing, while every GPT-6 Sol score so far is OpenAI's own. OpenAI lists no deprecation, and the $4/$20 rate is promotional through at least 21 November 2026. Test GPT-6 Sol before committing new workloads.

Entrée

$4 / 1M tokens

Sortie

$20 / 1M tokens

Contexte

Short $4/$20 · long $8/$30 per 1M · promo

GPT-Live-Transcribe

OpenAI

Recommandé

Le modèle parole-texte en streaming à faible latence d'OpenAI délivre des deltas de transcription en direct à 0,017 $/min avec une latence ajustable et une transcription contextuelle.

Entrée

$0.017 / minute

Sortie

N/A

Contexte

Per-minute realtime audio billing

GPT-Transcribe

OpenAI

Recommandé

Le dernier modèle speech-to-text en batch d'OpenAI obtient 3,31 % AA-WER à 0,0045 $/min — 25 % moins cher et mesurablement plus précis que son prédécesseur gpt-4o-transcribe. Il prend en charge les invites contextuelles, les suggestions de mots-clés et la détection de langue sur plus de 22 langues, ce qui en fait le meilleur choix par défaut pour la transcription asynchrone de fichiers sur la plateforme OpenAI. Réserve : pas de diarisation ni d'horodatage au niveau des mots ; utilisez gpt-4o-transcribe-diarize pour l'identification des locuteurs ou gpt-live-transcribe pour les besoins en temps réel. Idéal pour les développeurs qui construisent des pipelines de transcription de réunions, de traitement média et d'interfaces vocales sur l'API OpenAI.

Entrée

$0.0045 / minute

Sortie

N/A

Contexte

Per-minute audio billing

Claude Fable 5

Anthropic

Conditionnel

Anthropic's Mythos-class frontier model, the strongest publicly available Claude at its June 2026 launch, now listed as legacy behind Claude Fable 5.1. The verdict stays conditional on two constraints: API pricing of $10/$50 per 1M tokens, and a safety classifier that over-triggered on routine coding tasks at launch, routing them to an Opus 4.8 fallback. Subscribers reach Fable through Claude Max at 50% of weekly limits, or through usage credits on Pro. Anthropic's current lineup lists Fable 5.1 and Opus 5.5 for new work.

Entrée

$10 / 1M tokens

Sortie

$50 / 1M tokens

Contexte

1M ctx, 128K out; legacy; Max plans at 50% limits

Claude Fable 5.1

Anthropic

Conditionnel

Anthropic's new frontier flagship is the most capable model publicly available, and it largely fixes the safeguard over-triggering that made Fable 5 frustrating for routine coding. The verdict stays conditional on cost: independent measurement shows more spend per task than Fable 5 despite the cheaper cache reads, because the model thinks longer. Reach for it on demanding reasoning and long-horizon agentic work. For everyday coding, Claude Opus 5.5 at $4/$20 per 1M is Anthropic's recommended starting model and the better default.

Entrée

$10 / 1M tokens

Sortie

$50 / 1M tokens

Contexte

1M context, 128K output; cache $0.25/1M

Claude Haiku 5.5

Anthropic

Conditionnel

Claude Haiku 5.5 is a conditional pick: Anthropic's fastest and cheapest current Claude at $0.10 input and $0.50 output per 1M tokens, with a 1M context and adjustable effort. Prompts over 100K tokens cost five times as much, and every benchmark is Anthropic's own. Right for high-volume classification, extraction, routing and subagent work. Use Sonnet 5.5 for complex agentic coding, where Anthropic's own Terminal-Bench 4.0 numbers put it well ahead.

Entrée

$0.10 / 1M tokens

Sortie

$0.50 / 1M tokens

Contexte

1M ctx; prompts over 100K cost $0.50/$2.50

Claude Mythos 5.1

Anthropic

Conditionnel

Same weights as Claude Fable 5.1 at the same $10/$50 rate, with cyber and life-sciences safeguards loosened for vetted users. The catch is access: Mythos 5.1 ships only through Anthropic's trusted-access programs (Cyber Verification and Life Sciences Verification), so there is no self-serve API and no way to buy your way in. Conditional for the same reason as Mythos 5: elite capability behind a vetting gate. Everyone else should use Fable 5.1.

Entrée

$10 / 1M tokens

Sortie

$50 / 1M tokens

Contexte

1M context, $0.25/1M cache; vetted access

Claude Opus 5.5

Anthropic

Conditionnel

Claude Opus 5.5 is the model Anthropic now recommends starting with for most workloads, built for long-running agentic coding and knowledge work at rates 20% below Claude Opus 5. Conditional because every benchmark is Anthropic's own and migration is not drop-in: thinking cannot be turned off, forced tool use returns an error, and most cybersecurity tasks are rerouted to Opus 4.8. Right for Claude Opus 5 teams that can absorb those breaking changes.

Entrée

$4 / 1M tokens

Sortie

$20 / 1M tokens

Contexte

1M ctx; 128K out; cache $0.20; fast $8/$40

Claude Sonnet 4.6

Anthropic

Conditionnel

Superseded by Claude Sonnet 5 (shipped Jun 30, 2026), which Anthropic now lists as legacy behind Claude Sonnet 5.5 as the current Sonnet. Both list at $2/$10 per 1M tokens against Sonnet 4.6's $3/$15, since Anthropic cancelled the scheduled rise to $3/$15. Sonnet 5 also adds a 13.4-point Terminal-Bench 2.1 jump. Choose 4.6 only for pinned deployments or where Sonnet 5's ~30% higher token count and removed temperature/top_p/top_k are disqualifying.

Entrée

$3 / 1M tokens

Sortie

$15 / 1M tokens

Contexte

1M tokens

Claude Sonnet 5.5

Anthropic

Conditionnel

Conditional. Claude Sonnet 5.5 is Anthropic's default Sonnet model on the Anthropic API, on the same rate card as Sonnet 5, and Anthropic says it runs faster and costs less per task. Every benchmark is Anthropic's own, and five breaking changes hit code written for Sonnet 5. Right for teams ready to re-test agentic coding and knowledge work; hold off if you rely on forced tool choice or temperature control.

Entrée

$2 / 1M tokens

Sortie

$10 / 1M tokens

Contexte

1M ctx; batch $1/$5; cache read $0.10

DeepSeek V4.1 Flash

DeepSeek

Conditionnel

DeepSeek V4.1 Flash, the model now behind DeepSeek's deepseek-flash API, is a strong-value open-weight pick: MIT-licensed weights, image input, a 1M-token context and an Artificial Analysis score well above average for comparable models. Conditional because its standout coding-agent results are DeepSeek's own, Artificial Analysis measured it as very verbose, and peak hours double the rate. Right for cost-sensitive pipelines that can run off-peak or self-host; wrong for buyers who need independently verified agent performance before committing.

Entrée

$0.15 / 1M tokens

Sortie

$0.60 / 1M tokens

Contexte

1M tokens; 2x price at weekday peak hours

Conditional. Eleven v4 is ElevenLabs' long-form text-to-speech model (id eleven_v4), covering 90+ languages with multi-speaker dialogue and voice cloning, and accepting up to 10,000 characters per request. The deciding caveat is price: $0.022 per 1K characters is a 72%-off promo that ends Oct 12 against a $0.08 list rate, and the quality claims are ElevenLabs' own, untested by us. Suits teams producing content and long-form audio who budget at list price.

Entrée

$0.08 / 1K characters

Sortie

Billed per character (single rate)

Contexte

List $0.08 / 1K chars; 10K char limit

Conditional. Eleven v4 Turbo (id eleven_v4_turbo) is ElevenLabs' real-time speech model, stating a median inference latency of about 100 ms, excluding application and network latency, across 90+ languages via a Text to Dialogue WebSocket. The deciding caveat is that the latency figure is vendor-stated and untested by us, and $0.011 per 1K characters is a promo ending Oct 12 against a $0.04 list rate. Suits builders of voice agents and interactive characters.

Entrée

$0.04 / 1K characters

Sortie

Billed per character (single rate)

Contexte

List $0.04 / 1K chars; median latency ~100 ms

FLUX 3

Black Forest Labs

Conditionnel

FLUX 3 is Black Forest Labs' unified multimodal model for image, video, audio and robot-action prediction. Its Video component is now pay-as-you-go through BFL's dashboard and API, from $0.06 per second at Draft HD to $0.80 at UHD, with five seconds of Standard HD costing $0.85. Conditional because every preference result is BFL's own, Image and Audio are not yet released, and the open-weight FLUX 3 Dev is still listed as coming soon.

Entrée

$0.17 / sec (Standard HD text-to-video)

Sortie

Included in per-second video rate

Contexte

Video GA pay-as-you-go; $0.06-$0.80/s by res

Gemini 3.1 Pro

Google

Conditionnel

Le modèle Pro le plus performant de Google accessible au public, avec un contexte d'un million de tokens à un tarif compétitif. Génération précédente : Gemini 3.6 et 3.5 Flash le surpassent désormais en codage et en benchmarks agentiques à moindre coût. Restez-y si vous êtes dans l'écosystème Google Cloud/Vertex AI ; sinon, la gamme Flash ou les concurrents offrent davantage.

Entrée

$2 / 1M

Sortie

$12 / 1M

Contexte

1M tokens

Gemini 3.8 Flash Cyber

Google

Conditionnel

Gemini 3.8 Flash Cyber is Google's most capable cybersecurity model, a defensive specialist already in production at Chrome Security, Cloud Vulnerability Research and partner Wiz, with frontier vulnerability-discovery and patching results at a fraction of the leaders' rollout cost. The deciding caveat is access: it is gated to vetted defenders through the Fairwind Program, with no public API or published rate card. We rate it conditional rather than caution like the sibling Gemini 3.5 Flash Cyber because 3.8 ships published benchmark figures and named production deployments where 3.5 had neither, even though both remain gated. Right for organizations that can qualify as trusted defenders; everyone else cannot buy it today.

Entrée

N/A (no public rate card)

Sortie

N/A (no public rate card)

Contexte

Fairwind-gated; no public API

Gemini 3.8 Flash TTS

Google

Conditionnel

Conditional. Gemini 3.8 Flash TTS is Google's flagship text-to-speech model, available in the Gemini API with voice design, consented voice replication and SynthID watermarking on every clip. The deciding caveat is cost: today's paid rates are launch prices that double when the launch window closes, and its Hume AI leaderboard placements are Google-reported, not measured by us. Suits developers building expressive narration, character voices or dubbing who can budget for the post-launch rate.

Entrée

$0.50 / 1M (text)

Sortie

$9.00 / 1M (audio)

Contexte

Launch rate to 2026-12-31; $1/$18 from 2027

Gemini 3.8 Flash-Lite TTS

Google

Conditionnel

Conditional. Gemini 3.8 Flash-Lite TTS is Google's cost tier for high-volume speech: stable in the Gemini API, built into Google Vids, and billed at the same text input rate as Flash TTS with a lower audio output rate. Google ranks it second only to Flash TTS on Hume AI's Overall Quality Index, but that ranking is Google-reported and its launch rates are temporary, doubling once the promotion ends. Suits voice-agent cascades and bulk narration where cost per clip matters more than acting range.

Entrée

$0.50 / 1M (text)

Sortie

$6.00 / 1M (audio)

Contexte

Launch rate to 2026-12-31; $1/$12 from 2027

Gemini 3.8 Live

Google

Conditionnel

Gemini 3.8 Live is Google's default Live API model for low-latency voice agents that keep talking while tool calls finish in the background. Conditional because Google bills Gemini 3.8 Live Extended Thinking at the identical rate, and Artificial Analysis scores that sibling higher on its Speech to Speech Index, so you pay the same for less reasoning. Pick 3.8 Live only when turn latency matters more than reasoning depth: developers shipping real-time voice experiences on the Gemini API.

Entrée

$0.75 / 1M (text)

Sortie

$4.50 / 1M (text)

Contexte

131K in, 65K out; audio $3/$12 per 1M

GLM-5.2

Z.ai

Conditionnel

Meilleur modèle de codage open weight avec un contexte de 1M tokens et licence MIT. Il bat GPT-5.5 sur plusieurs benchmarks pour une fraction du coût, mais reste derrière Claude Opus 4.8 sur les tâches à horizon long les plus difficiles.

Entrée

$1.40 / 1M tokens

Sortie

$4.40 / 1M tokens

Contexte

1M tokens

GLM-5.3

Z.ai

Conditionnel

GLM-5.3 is Z.ai's open-weights coding and cyber-defense model on the GLM-5.2 base, with every gain from post-training. Pricing is published at $1.40 input / $4.40 output per 1M, the same as GLM-5.2, and the weights are now on Hugging Face. It stays conditional because its headline coding and cyber scores come from Z.ai's own evaluations: reach for defensive security and agentic automation, but audit it yourself before standardizing.

Entrée

$1.40 / 1M tokens

Sortie

$4.40 / 1M tokens

Contexte

1M ctx; Flash $0.15/$0.50, launch promo ended

GPT-4o (Retired from ChatGPT, API-only)

OpenAI

Conditionnel

GPT-4o (May 2024) has been superseded by the GPT-5.x family and delisted from OpenAI's flagship API pricing page. It remains available through the API but is no longer a current-generation model. Use only where migration to GPT-5.x is blocked; otherwise, GPT-5.5 or GPT-5.4 offer better performance at comparable cost.

Entrée

$2.50 / 1M tokens

Sortie

$10 / 1M tokens

Contexte

128K tokens

GPT-5.5-Cyber

OpenAI

Conditionnel

85,6 % CyberGym — meilleur score mono-modèle — mais accès restreint aux défenseurs habilités uniquement.

Entrée

N/A

Sortie

N/A

Contexte

Trusted Access for Cyber; not commercial API

GPT-5.6 Luna

OpenAI

Conditionnel

GPT-5.6 Luna is OpenAI's previous-generation budget model at $0.20 input and $1.20 output per 1M tokens short context, generally available through the API and Codex. Its successor GPT-6 Luna, released 22 September 2026, lists at $0.10/$0.50, so Luna is no longer the low-cost default for new OpenAI work. It remains conditional because benchmarks are vendor-reported and it lacks activation classifiers; OpenAI lists no deprecation, so existing pipelines need not migrate yet.

Entrée

$0.20 / 1M tokens

Sortie

$1.20 / 1M tokens

Contexte

Short $0.20/$1.20 · long $0.40/$1.80 per 1M

GPT-5.6 Terra

OpenAI

Conditionnel

GPT-5.6 Terra is OpenAI's mid-tier workhorse: GPT-5.5-class capability on OpenAI's own claim, at $2.00 input and $12.00 output per 1M tokens at short context ($4.00 / $18.00 long) after the 20% price cut of 30 July 2026. It is generally available via the API and Codex. Remains conditional: all benchmarks are vendor-reported without independent reproduction, and Terra is not selectable in the standard ChatGPT model picker.

Entrée

$2.00 / 1M tokens

Sortie

$12.00 / 1M tokens

Contexte

Short $2.00/$12.00 · long $4.00/$18.00 per 1M

GPT-5.6-Cyber

OpenAI

Conditionnel

GPT-5.6-Cyber est le modèle de cybersécurité spécialement entraîné d'OpenAI, construit sur GPT-5.6 Sol et disponible exclusivement via le programme contrôlé Daybreak Red. C'est le premier modèle à atteindre le seuil de capacité cyber 'High' d'OpenAI, avec des performances en chaînes d'exploit et zero-day qui dépassent largement les modèles généralistes.

Entrée

$12.50 / 1M tokens

Sortie

$75 / 1M tokens

Contexte

400K tokens

GPT-6 Astra

OpenAI

Conditionnel

Astra ships at $10/$50 per 1M tokens short context, $20/$75 long context, with a 1,050,000-token context window and 128K max output. On OpenAI's own reported benchmarks it leads on computer use, terminal work, long context and cyber. We previously placed it third on the Artificial Analysis Intelligence Index at 61.2; we could not source that figure and have withdrawn it, and Artificial Analysis publishes no Astra score, so we have no vendor-neutral general-intelligence ranking for it. Conditional on two constraints: Fast mode is unavailable under EU data residency, and OpenAI disclosed decreased monitorability versus GPT-5.6 Sol. Route to it by task; do not make it your default.

Entrée

$10 / 1M tokens

Sortie

$50 / 1M tokens

Contexte

1.05M tokens; short-ctx standard rate

GPT-6 Luna

OpenAI

Conditionnel

GPT-6 Luna is OpenAI's new budget model and the cheapest way into GPT-6, at half GPT-5.6 Luna's input rate, with a 1.05M-token context and text and image input. Conditional because every published result is OpenAI's own and prompts over 272K input tokens bill at higher rates. GPT-5.6 Luna stays on the price list with no deprecation notice, so migration is optional. Right for teams running high-volume, everyday API work at scale.

Entrée

$0.10 / 1M tokens

Sortie

$0.50 / 1M tokens

Contexte

1.05M ctx; over 272K $0.20/$0.75; cache $0.01

GPT-6 Sol

OpenAI

Conditionnel

GPT-6 Sol is the lower-priced companion to GPT-6 Astra and succeeds GPT-5.6 Sol at half its per-token rate, with a 1.05M-token context and 128K output. Conditional because every published score is OpenAI's own and prompts over 272K input tokens bill at higher rates. GPT-5.6 Sol stays on the price list with no deprecation notice, so nothing forces a migration. Right for teams routing hard coding and agent work to OpenAI.

Entrée

$2 / 1M tokens

Sortie

$10 / 1M tokens

Contexte

1.05M ctx; over 272K $4/$15; cache $0.20

GPT-6.1 Sol

OpenAI

Conditionnel

Conditional on evidence: GPT-6.1 Sol is OpenAI's model for complex coding, computer use and professional work, which OpenAI itself describes as near-Astra performance at a lower cost. We could not read the launch post, so no benchmark is stored and the quality claim is OpenAI's alone. Prompts past 272K input tokens bill at higher rates, and reasoning cannot be set to none or minimal. Suits teams building agents on the Responses API who will run their own evals first.

Entrée

$2 / 1M tokens

Sortie

$10 / 1M tokens

Contexte

1.05M ctx; over 272K $4/$15; cache $0.10

GPT-Live-1

OpenAI

Conditionnel

GPT-Live-1 is OpenAI's full-duplex voice model: it listens and speaks at the same time and hands reasoning and tool use to a backend model. Since 10 September 2026 it is generally available in the API on the Live endpoint at $0.05 per minute, with backend model and tool usage billed separately. Conditional because its benchmarks are OpenAI's own, language quality is uneven, and real cost per call depends on the backend model you pair it with.

Entrée

$0.05 / minute (voice session)

Sortie

Included in per-minute session rate

Contexte

Live API; backend model + tools billed separately

GPT-Live-1 mini

OpenAI

Conditionnel

La variante gratuite de GPT-Live-1 — même architecture duplex intégral dans un package plus petit et plus rapide.

Entrée

Free (ChatGPT)

Sortie

Free (ChatGPT)

Contexte

ChatGPT Voice only

Grok 4.5

SpaceXAI

Conditionnel

Competitively priced coding and agentic model at $2/$6 per 1M tokens up to 200K tokens of context, doubling to $4/$12 above that, with a 500K window. xAI's published benchmarks place it mid-pack behind Fable 5 and GPT-5.5, though it leads Opus 4.8 on DeepSWE and Terminal-Bench 2.1. The real story is cost efficiency: 4.2x more token-efficient than Opus 4.8 on SWE-Bench Pro. Best for teams needing strong coding at a fraction of frontier pricing; xAI's docs list only US regions.

Entrée

$2 / 1M tokens

Sortie

$6 / 1M tokens

Contexte

500K ctx; 200K+ $4/$12; cache $0.30 (200K+ $0.60)

Grok 4.6

SpaceXAI

Conditionnel

Grok 4.6 is SpaceXAI's frontier model for coding and agentic work at $2 input and $6 output per 1M tokens under 200K, with a 500K-token context; prompts of 200K or more bill at $4/$12. At launch it scored 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol, but that is still one independent evaluator. Grok 4.7 now lists at the same price, so compare the two before committing new work.

Entrée

$2.00 / 1M tokens

Sortie

$6.00 / 1M tokens

Contexte

500K ctx; 200K+ $4/$12; cache $0.50

Grok 4.7

SpaceXAI

Conditionnel

Grok 4.7 keeps Grok 4.6's price and speed, and xAI says it works longer on hard tasks and checks its own work more carefully. Conditional because every published score is xAI's own, prompts of 200K tokens or more bill at double the headline rate, and the Batch API is not supported. Right for coding teams already working in Cursor or Grok Build.

Entrée

$2.00 / 1M tokens

Sortie

$6.00 / 1M tokens

Contexte

500K ctx; 200K+ $4/$12; fast variant 2x price

Grok Voice 3.0

SpaceXAI

Conditionnel

Grok Voice 3.0 is SpaceXAI's speech-to-speech model for real-time voice agents, priced at $0.08 per minute, the same as Grok Voice 2.0. Artificial Analysis shows the highest Speech-to-Speech Index score, 86.9, with 0.81 seconds to first audio, but xAI's other benchmark and customer figures are its own. Conditional because we have not tested it and the grok-voice-latest alias moved to it automatically. It suits teams building phone or in-app voice agents on a managed model.

Entrée

$0.08 / min of audio (speech-to-speech)

Sortie

Included in per-minute audio rate

Contexte

Voice Agent API; $4.80 / hr

Grok Voice Think Fast 2.0

SpaceXAI

Conditionnel

Le modèle speech-to-speech de nouvelle génération de SpaceXAI, doté d'une architecture de raisonnement parallèle qui le rend nettement plus intelligent que les modèles vocaux conventionnels.

Entrée

$0.08 / min of audio (speech-to-speech)

Sortie

Included in speech-to-speech rate

Contexte

API-only

Grok Voice Transcribe 2.0

SpaceXAI

Conditionnel

Grok Voice Transcribe 2.0 is SpaceXAI's speech-to-text model at $0.10 per hour of audio in batch, with speaker labels, word timestamps and key terms included. Conditional because every accuracy figure is SpaceXAI's own: it says the model is twice as accurate as 1.0 and first among 32 streaming models on Artificial Analysis, and we could not check that leaderboard. Right for voice-agent and call-transcription developers who will test it on their own audio; we have not tested it.

Entrée

$0.10 / hour of audio (batch)

Sortie

Included in per-hour audio rate

Contexte

Streaming $0.20 / hour of audio

Hy3

Tencent

Conditionnel

Apache 2.0 open-weight MoE agentic model (295B total, 21B active) with strong tool-calling reliability and anti-hallucination behavior, but coding trails GLM-5.2 and self-hosting demands 8+ GPUs. The free OpenRouter tier that ran through Jul 21 is gone; OpenRouter now lists six providers at $0.13-$0.20 input and $0.53-$0.80 output per 1M tokens. Best for self-hosted coding agents and long-context document work where permissive licensing matters more than absolute benchmark scores.

Entrée

$0.20 / 1M

Sortie

$0.80 / 1M

Contexte

256K tokens; OpenRouter providers $0.13-$0.20 in / $0.53-$0.80 out; free tier ended

Inkling

Thinking Machines Lab

Conditionnel

Inkling est la meilleure raison à ce jour de prendre l'IA multimodale open-weight au sérieux. Il offre un raisonnement et un codage compétitifs avec une sécurité open-weight de premier ordre et un effort de réflexion contrôlable, le tout sous licence Apache 2.0. Mais Thinking Machines est honnête : ce n'est pas le plus performant dans l'absolu — la factualité est en retrait (SimpleQA 43,9 %) et l'auto-hébergement exige du matériel sérieux (2 To de VRAM). Conditionnel : une base convaincante pour les organisations qui ont besoin de posséder et de fine-tuner un modèle multimodal, pas pour les équipes qui veulent la performance maximale sur étagère.

Entrée

$1.87 / 1M tokens

Sortie

$4.68 / 1M tokens

Contexte

Up to 1M tokens native; Tinker 64K and 256K tiers; rates at 50% limited-time discount (list $3.74 / $9.36)

Kimi K3

Moonshot AI

Conditionnel

Moonshot AI's open-weight MoE model (Kimi K3 License, weights on Hugging Face) offers a 1M-token context at $3/$15 per 1M tokens and holds #1 on Arena WebDev (preliminary), but still trails Fable 5 and GPT-5.6 Sol on precise execution, and serving the weights needs 64+ accelerators. The rating stays conditional on legal risk: on July 22, 2026 the White House OSTP accused Moonshot of distilling Fable 5 and Treasury threatened sanctions. We have not re-verified that dispute's status this run.

Entrée

$3.00 / 1M

Sortie

$15.00 / 1M

Contexte

1M ctx, $0.30 cached. Weights live (Jul 27).

Kolibri-1 is Aleph Alpha's English-German mixture-of-experts model with Apache 2.0 weights on Hugging Face, and a credible open-weight option for teams that must run German-language workloads on their own hardware. The catch is operations: the FP8 weights need about 78 GB of GPU memory, the pages we read name no hosted API or price, and every benchmark comes from Aleph Alpha's own harnesses. Right for public-sector and industrial teams that can self-host. Skip it for hosted-API needs or other languages.

Entrée

Free (open weights)

Sortie

Free (open weights)

Contexte

262K native, 1M extrapolated; ~78 GB FP8

Laguna S 2.1

poolside

Conditionnel

Laguna S 2.1 est le premier modèle de coding open-weight occidental crédible en 11 mois — un MoE de 118B qui n'active que 8B de paramètres par token tout en obtenant 70,2 % sur Terminal-Bench 2.1, battant des modèles 10 fois sa taille. Mais il n'est pas au niveau frontière — GPT-5.6 Sol et Kimi K3 sont loin devant — et les limitations connues incluent le surapprentissage de harnais et la déformation JSON. Idéal pour le coding agentique auto-hébergé lorsque la souveraineté des données prime sur le leadership absolu des benchmarks.

Entrée

$0.10 / 1M

Sortie

$0.20 / 1M

Contexte

OpenMDW-1.1 license, 1M context

Llama 4 Maverick

Meta

Conditionnel

Le modèle open source de référence pour les équipes qui ont besoin d'auto-hébergement, de souveraineté des données, ou qui cherchent à éviter l'enfermement propriétaire des API.

Entrée

Free (self-hosted) or $0.20 / 1M tokens (hosted)

Sortie

Free (self-hosted) or $0.60 / 1M tokens (hosted)

Contexte

1M tokens

MAI-Code-1-Flash

Microsoft

Conditionnel

MAI-Code-1-Flash was Microsoft's first in-house coding model, a sparse MoE priced at $0.75/$4.50 per 1M tokens. GitHub retired it from Copilot on 10 September 2026 and names MAI-Code-1.1-Flash, at $0.20/$1.20, as the replacement; it no longer appears in Copilot's supported-models or pricing tables. Copilot users should switch to 1.1-Flash, which is cheaper and scores higher on Terminal-Bench 2.1.

Entrée

$0.75 / 1M tokens

Sortie

$4.50 / 1M tokens

Contexte

Retired from Copilot Sep 10 2026; 256K ctx

MAI-Code-1.1-Flash

Microsoft

Conditionnel

MAI-Code-1.1-Flash est le modèle de codage small-tier actualisé de Microsoft pour Copilot — un quart du coût et 22 % meilleur sur Terminal-Bench 2.1 que son prédécesseur de juin, désormais avec la vision native. Il est conçu pour le travail itératif à fort volume plutôt que le raisonnement de frontière, ce qui en fait le modèle de codage sérieux le moins cher de Copilot. Tournez-vous vers lui pour les complétions quotidiennes, les refactorisations et le débogage assisté par image ; utilisez un modèle phare pour l'architecture complexe.

Entrée

$0.20 / 1M

Sortie

$1.20 / 1M

Contexte

256K context

MAI-Cyber-1-Flash

Microsoft

Conditionnel

Microsoft's first in-house cybersecurity model, built on the MAI-Thinking-1 family and embedded inside the MDASH multi-agent harness, handles about 90% of security workloads at roughly half the cost of prior configurations and reserves GPT-5.4 for the hardest 10%. Microsoft reports 96% on CyberGym by the any-crash metric and 86.3% by the stricter final-submission metric, both from a tuned full-system configuration rather than an independent model-versus-model test. Available only through Azure AI Foundry with vetted enterprise access; no standalone pricing or public API.

Entrée

N/A (MDASH SCU consumption)

Sortie

N/A (MDASH SCU consumption)

Contexte

Azure AI Foundry vetted access

Conditional. Mercury 2.5 is Inception's diffusion LLM, and raw output speed is the reason to pick it: Artificial Analysis ranks it among the fastest models it measures, and it ships tunable reasoning, parallel tool calls and structured outputs. The deciding caveat is quality: Artificial Analysis places its intelligence just below the median, and Inception's parity claims name no benchmark. Suits latency-bound voice agents, RAG pipelines and coding subagents, not hard reasoning.

Entrée

$0.20 / 1M tokens

Sortie

$0.75 / 1M tokens

Contexte

260K ctx; 80% off at launch: $0.04/$0.15

Mercury 3 is Inception's diffusion LLM with a 1M-token context at a list price of $0.20 input and $0.70 output per 1M tokens, and speed is the reason to pick it. The deciding caveat is that its quality claims come from Inception and its partners, and the headline comparisons name no benchmark. Suits voice, chat, search and browser agents where latency sets the experience, not hard reasoning.

Entrée

$0.20 / 1M tokens

Sortie

$0.70 / 1M tokens

Contexte

1M ctx; list rate shown; 50% off at launch ($0.10 / $0.35), no end date stated, so plan on the list rate

MiniMax M3.1 Flash is a 1M-context multimodal coding and agent model at $0.15 input and $0.60 output per 1M tokens, and MiniMax says its agentic usage cost is 70% lower than its predecessor's. The deciding caveat is evidence: the launch post's scores are MiniMax's own, it names no comparison model, and the weights are promised only for later. Worth a trial for cost-sensitive agent and computer-use workloads, not yet for anyone who needs self-hosting or independent scores.

Entrée

$0.15 / 1M tokens

Sortie

$0.60 / 1M tokens

Contexte

1M context; cached input $0.015 / 1M

Mistral Large 4

mistral

Conditionnel

Conditional: Mistral Large 4 is a public preview, not a finished release, since Mistral says its reinforcement learning run is still in flight and promises weights "by the end of the month". It is a mixture-of-experts model of about 1 trillion parameters, 52B active, with a 1M context on its docs page and a $1.36 input, $4.18 output list price per 1M tokens. Most benchmarks are Mistral's own. Trial it now if you want an open-weight-bound model, but wait for weights and licence before committing.

Entrée

$1.36 / 1M tokens

Sortie

$4.18 / 1M tokens

Contexte

1M ctx; preview; sale $0.68/$2.09 on docs

Muse Glimmer

Meta

Conditionnel

Muse Glimmer est le modèle agentic open-weight de 30B de Meta, distillé à partir du modèle frontier fermé Muse Spark et publié sous Apache 2.0. Avec une quantification 4-bit, il exécute des appels d'outils multi-étapes, du raisonnement multimodal et de la récupération d'erreurs sur un seul GPU grand public 24 Go sans dépendances cloud.

Entrée

Free (open weights)

Sortie

Free (open weights)

Contexte

131K tokens

Muse Spark 1.1

Meta

Conditionnel

Le premier modèle d'IA payant de Meta et sa première salve sur le marché du coding commercial — 1,25 $/4,25 $ par 1M tokens avec 1M de contexte, orchestration native de sous-agents, support MCP/outils et capacité computer-use à environ 4× sous les tarifs d'Anthropic/OpenAI. Domine les benchmarks d'utilisation d'outils mais accuse un retard de 7 points sur Opus 4.8 en coding pur sur SWE-Bench Pro. Poids fermés, préversion publique réservée aux États-Unis. Idéal pour les équipes privilégiant l'orchestration d'outils à bas coût.

Entrée

$1.25 / 1M tokens

Sortie

$4.25 / 1M tokens

Contexte

1M tokens

Muse Spark 1.2

Meta

Conditionnel

La mise à jour du modèle de Meta axée sur le codage, co-entraîné avec le harnais Muse Code — 82,9 % sur Terminal-Bench 2.1 et 59,3 % sur DeepSWE 1.1, deuxième seulement derrière Claude Opus 5 sur le premier. Tarification standard inchangée à 1,25 $ en entrée / 4,25 $ en sortie par million de tokens.

Entrée

$1.25 / 1M tokens

Sortie

$4.25 / 1M tokens

Contexte

1M tokens

Muse Spark 1.3

Meta

Conditionnel

Muse Spark 1.3 is Meta's strongest coding and long-context model yet, and an easy drop-in for anyone already on 1.2: the standard token price is unchanged while software-engineering and million-token retrieval results move sharply up. The deciding caveat is that the headline agent numbers come from a max reasoning configuration Meta has not released, while the shipping tier runs slower and costs more per task. Right for cost-sensitive coding teams inside Muse Code; wrong for anyone who needs independently reproduced benchmarks.

Entrée

$1.25 / 1M tokens

Sortie

$4.25 / 1M tokens

Contexte

1M ctx; contributor tier $0.10/$0.20

Muse Spark Open

Meta

Conditionnel

Muse Spark Open is Meta's open-weight release of Muse Spark, an 816B mixture-of-experts model with 42B active parameters and a 1M-token context, available as downloads and from seven hosted providers. Conditional: the Muse Model License has limited commercial terms we did not read, the Meta pages we read give no hosted price, and Meta says its full evaluation report is still to come. It suits teams that want to self-host or shop between hosts, not buyers who need audited terms or independent benchmarks.

Entrée

Free (open weights)

Sortie

Free (open weights)

Contexte

1M ctx; Muse Model License (limited commercial terms); no hosted price read

Muse Wonder

Meta

Conditionnel

Muse Wonder is Meta's most capable model on its API, but it is still a preview and we could not read a single benchmark score behind the launch claims. Its list price is 3.2x Muse Spark 1.3 on input and 4.7x on output, and Meta says the full safety report is still to come. It suits teams already on Muse Spark who want to test it on long-horizon agent work, not buyers who need independently checked results.

Entrée

$4.00 / 1M tokens

Sortie

$20.00 / 1M tokens

Contexte

1M ctx; preview; cached $0.20

Mythos 5 (Claude Mythos 5)

Anthropic

Conditionnel

Le modèle de cybersécurité le plus puissant au monde (83,8% CyberGym), construit sur les poids de Fable 5 avec les classificateurs de sécurité retirés. L'accès a été suspendu après la directive d'exportation du 12 juin ; partiellement restauré le 26 juin pour ~100 organisations américaines désignées d'infrastructure critique (Project Glasswing). Le verdict reste CONDITIONNEL : la capacité est d'élite, mais l'accès est verrouillé par désignation fédérale. Pas une API publique : utilisable uniquement si votre organisation figure sur la liste du Département du Commerce.

Entrée

$10 / 1M tokens

Sortie

$50 / 1M tokens

Contexte

1M tokens (restricted access)

Nemotron 3.5 Super VL is NVIDIA's open-weight 120B vision-language model (12B active) under the OpenMDW-1.1 licence, with a 1M-token context and text, image and video input. The deciding caveat is hardware and evidence: NVIDIA recommends 4x H100 or 2x GB200 for the BF16 build, and the card's scores are NVIDIA's own runs. Suits teams that can host a large model on their own GPUs and want open weights for document and visual work.

Entrée

Not stated

Sortie

Not stated

Contexte

Open weights under OpenMDW-1.1; 1M ctx; hosted NIM and OpenRouter prices not verified

Ornith-1.0

DeepReinforce

Conditionnel

Une famille de modèles open source agentic coding sous licence MIT qui atteint des benchmarks compétitifs avec les modèles frontier — battant Claude Opus 4.7 à la fois sur Terminal-Bench 2.1 et SWE-Bench Verified — mais qui nécessite un autohébergement avec des ressources GPU significatives et provient d'un nouveau labo au track record limité.

Entrée

Free (MIT license)

Sortie

Free (MIT license)

Contexte

256K tokens

Palmyra X6

Writer

Conditionnel

Palmyra X6 est le modèle phare agentique de Writer pour les workflows marketing et de revenus — post-entraîné sur le GLM-5.2 de Z.ai (poids ouverts), à $2/$8 par 1M de tokens, une fraction du tarif des modèles de pointe. Chaque chiffre clé provient de Writer, et la provenance GLM-5.2 soulève une vraie question d'approvisionnement. Adoptez-le au sein de Writer pour le travail d'agents GTM sensible au coût ; cherchez ailleurs pour une capacité générale benchmarkée de façon indépendante.

Entrée

$2 / 1M

Sortie

$8 / 1M

Contexte

1M context, 8K max output

Qwen-Image-2.1

Alibaba

Conditionnel

Qwen-Image-2.1 is Alibaba's open-weights model for both text-to-image generation and editing, with native transparent RGBA output and multi-image composition from up to 10 references. Conditional because the weights ship under the Qwen Research License, which allows non-commercial use only, and we found no readable benchmark scores. Right for researchers and hobbyists running image generation and editing locally through Diffusers.

Entrée

Free (non-commercial weights)

Sortie

Free (non-commercial weights)

Contexte

Qwen Research License; non-commercial

Qwen3.8-27B

Alibaba

Conditionnel

Qwen3.8-27B est le modèle open-weight à 27 milliards de paramètres d'Alibaba, désormais disponible sous licence Apache 2.0 avec 262K de contexte natif et un déploiement sur une seule GPU. C'est le compagnon auto-hébergeable du flagship Qwen3.8-Max à 2,4T. Les benchmarks restent des chiffres d'Alibaba, d'où le verdict 'conditional' : un modèle réel et exécutable dont les gains de codage revendiqués par le fournisseur attendent encore une reproduction indépendante.

Entrée

Free (open weights — self-hosted)

Sortie

Free (open weights — self-hosted)

Contexte

262K native ctx, 1M extended; Apache 2.0

Qwen3.8-Omni-Flash

Alibaba

Conditionnel

Qwen3.8-Omni-Flash takes text, images, audio and video in one call with a 1M-token context, and Alibaba says it cut the per-hour price of audio input by more than 98% against the previous version. Conditional because every benchmark we found is Alibaba's own and the model answers in text only; spoken replies need a separate model, such as the Realtime variant, at higher rates. Right for teams building audio-visual agents on Alibaba Cloud.

Entrée

$0.15 / 1M

Sortie

$0.47 / 1M (text)

Contexte

1M ctx; cache hit $0.016/1M; International

Sakana Fugu

Sakana AI

Conditionnel

Orchestrateur multi-agents qui égale les modèles de pointe sur les principaux benchmarks en routant à travers un pool d'agents experts — mais le routage opaque et la dépendance à la plateforme limitent la transparence.

Entrée

$5 / 1M tokens

Sortie

$30 / 1M tokens

Contexte

272K tokens (>272K: $10/$45 per 1M)

SWE-2

Cognition

Conditionnel

SWE-2 is Cognition's coding model for Devin, post-trained from Kimi K3, and Cognition claims it lands within one point of Fable 5.1 on FrontierCode 1.1 Main at 64% lower cost. Conditional because every SWE-2 score is vendor-reported, the same table shows it far behind Fable 5.1 and GPT-6 Astra on Terminal-Bench 4, and it runs only inside Devin. Right for Devin Desktop and CLI users; wrong for anyone who needs an API model or independently verified results.

Entrée

N/A (no standalone API)

Sortie

N/A (no standalone API)

Contexte

Devin only; Enterprise $0.75 in/$3.75 out per 1M

Ternary Bonsai 2 27B

PrismML

Conditionnel

Ternary Bonsai 2 27B squeezes Qwen3.8 27B into about 5.9GB at 1.76 effective bits, so a 262K-context text-and-image model fits on a Mac or a single consumer GPU under Apache 2.0. Conditional because the quality figures are PrismML's own, its blog and model card report different aggregate scores, and the files need PrismML's custom kernels: stock llama.cpp will not run them. Right for local-inference users short on memory.

Entrée

Free (open weights)

Sortie

Free (open weights)

Contexte

262K ctx; self-hosted; PrismML kernels required

Gemini 3.5 Flash Cyber

Google

Prudence

Gemini 3.5 Flash Cyber est le premier modèle spécialisé en cybersécurité de Google, fine-tuné à partir de 3.5 Flash pour trouver, valider et corriger des vulnérabilités. Au sein de l'architecture multi-agent de CodeMender, il atteint des performances compétitives sur CyberGym face à des modèles bien plus gros comme Mythos, pour une fraction du coût. Mais les restrictions sont sévères : il fonctionne exclusivement dans CodeMender, disponible uniquement pour les gouvernements et partenaires de confiance via un pilote à accès limité, sans API publique, sans chiffres de benchmark publiés et sans calendrier pour une diffusion plus large.

Entrée

N/A

Sortie

N/A

Contexte

CodeMender-only, government-gated pilot

Gemini 3.5 Pro

Google

pending

Le modèle frontière pré-GA de Google, désormais à son quatrième report et « des mois de retard » selon Bloomberg. Le contexte de 2M tokens, le pipeline multimodal natif et le raisonnement Deep Think restent des capacités réelles, mais aucun calendrier de disponibilité générale confirmé n'existe. La capacité de codage est la lacune principale.

Entrée

Not yet published

Sortie

Not yet published

Contexte

Est. $5-15/$30-60 per 1M. 2M ctx. Not GA.

Gemini 4

Google

pending

Le prochain modèle de Google DeepMind, confirmé. Le pré-entraînement a débuté le 21 juillet 2026. Aucun benchmark, aucun tarif, aucune spécification, aucune date de GA.

Entrée

$2 / 1M tokens (through Nov 30)

Sortie

$10 / 1M tokens (through Nov 30)

Contexte

$2/$10 to Nov 30, then $4/$20; paid tier only

Qwen3.8-Max

Alibaba

pending

Flagship multimodal MoE 2,4T paramètres. Pas encore GA : aucun tarif par token pour l'API, aucun tableau de benchmarks, aucun poids ouvert publié. Entrée à 6 $/mois via Token Plan Lite.

Entrée

$2.00 / 1M tokens

Sortie

$6.00 / 1M tokens

Contexte

Model Studio (Singapore); one tier up to 1M tokens