Same Name, Different Deal: DeepSeek's Swapped Model, Sonnet 5's Permanent Price and a Reported Pro Pause
Every story we published this week changed what an existing name buys you: the deepseek-v4-flash id now serves V4.1-Flash at $0.15/$0.60 per 1M off-peak, Claude Sonnet 5's $2/$10 launch price is now its standard rate, and GLM-5.3-Flash's $0.075/$0.25 launch rate gave way to $0.15/$0.50. TechCrunch reports OpenAI has paused new $200 ChatGPT Pro sign-ups, which leaves the API and Plus as the ways to reach Astra.
125 announcements scanned · 4 mattered · 0 verdicts changed
Verdicts modifiés
Claude Sonnet 5 (basis change, rating held)
The rating did not move, so this is not counted in the week's verdict changes. The basis did: our entry described $2/$10 per 1M as introductory pricing that expired on 31 August 2026. Anthropic cancelled the scheduled increase to $3/$15, so $2/$10 is now the standard rate, 40 percent of the $5/$25 Opus rate. Announced September 14 by claude-sonnet-5-price-cut-made-permanent.
GLM-5.3 (basis change, rating held)
The rating did not move, so this is not counted in the week's verdict changes. The summary was corrected: it said per-token pricing was unpublished, but Z.ai lists GLM-5.3 at $1.40/$4.40 per 1M, and the entry had stored the expired GLM-5.3-Flash launch promotion as the list rate. GLM-5.3 stays conditional because its headline coding and cyber scores come from Z.ai's own evaluations. Announced September 14 by glm-5-3-flash-launch-discount-ends.
À traiter avant le prochain numéro
Re-run your evals on anything that calls deepseek-v4-flash
DeepSeek released V4.1-Flash on 10 September and retired V4 Flash the same day. The old model id now routes to V4.1-Flash and bills at the new Flash price of $0.15 in and $0.60 out per 1M tokens off-peak. The id is the same, but the model behind it is not.
Détails et étapesBudget GLM-5.3-Flash at $0.15/$0.50, not the launch rate
The $0.075/$0.25 launch rate was a 50 percent promotion that ended on 9 September 2026. Full-size GLM-5.3 stays at $1.40/$4.40 per 1M, and our verdict stays conditional.
Détails et étapesReach Astra through the API or Plus while Pro sign-ups are paused
TechCrunch reports that OpenAI's Thibault Sottiaux said on X that new sign-ups to the $200-a-month ChatGPT Pro plan are paused because Pro puts the most strain on OpenAI's systems. The API, Go and Plus stay open, and no end date was given.
Détails et étapesSur notre radar
DeepSeek still lists V4 Pro separately after September 14 — This updates the Week 37 radar item on DeepSeek's two conflicting pages. DeepSeek's pricing page, read on September 23, still lists deepseek-v4-pro (DeepSeek-V4-Pro-0813) at $0.66 input and $1.98 output per 1M tokens off-peak, while the September 10 launch post still says V4 Pro requests route to V4.1-Flash from 04:00 UTC on September 14. DeepSeek has not reconciled the two, so we have not confirmed which model answers deepseek-v4-pro calls. Our V4 Pro entry stays recommended.
DefenseScoop: the DoD is moving its classified AI workloads off Anthropic by October — DefenseScoop, September 11 (body fetched): Emil Michael says "about 90% has transitioned". It is a government-customer move and does not change what a commercial reader can buy. Held until a ruling or designation affects non-government availability.
Grok 4.7 delayed, per Musk's posts on X — CryptoBriefing, September 11 (body fetched), sourced only to Musk's posts on X. Held until xAI publishes a release post or a live API model id.
Researchers link OpenAI agents to the May RubyGems campaign — Reported by The Hacker News and The Verge on September 12 (the researchers' own site was not fetched). It extends the agent-swarm story from September 7. Held until the primary findings are fetched.
OpenAI's dispute with mathematicians escalates to an open letter — TechCrunch, September 11 (RSS summary only): 25 mathematicians signed. A dispute over training-data provenance with no product consequence. Held until a legal filing or an OpenAI disclosure.
Microsoft opened a six-week review of draft rules for its MAI models — The Verge, September 14 (feed content), alongside a 37-page code of conduct. It could affect MAI model behaviour but changes no product yet. Held until final rules change behaviour or access.
OpenAI Astra for Law — OpenAI News feed, September 17, description only. The post body could not be read. Held until a page shows a new model id, price or access tier.
TypeSafe's System One models and Jev are in early access — TypeSafe, September 15 (fetched): waitlist only, so readers cannot use it yet. The vendor quotes $0.042 per 1M input tokens, with output free. Held until general availability or self-serve API signup.
Meta Muse adds calling — TechCrunch, September 17 (RSS summary only). A feature added to a directory tool. Held until a Meta primary describes the feature and which plans get it.
Mandiant: an attacker hijacked an active AI coding-assistant session — The Hacker News, September 16 (body fetched): Shai-Hulud spread across about 100 repos. Mandiant does not name the assistant or disclose how the session was taken over. Held until either is named.
Held on a headline or search summary only — Not yet read, so none of these is written up: Gemini 4 Pro reportedly pushed toward October, a reported Sanders and Casar bill to outlaw superintelligent AI, a reported OpenAI Pentagon contract defined by refusal rates, reported code suggesting Siri can swap in Claude or ChatGPT, Claude Skills API reportedly out of beta, reported launches of Sakana Fugu Ultra v2.0 and Tencent Hy4 Preview, a reported Mistral x Mozilla partnership, a reported investigation into AI security-eval firm Irregular, a claim that DeepSeek V4.1 Flash is the best hacking model, reports that iOS 27 has no Apple Intelligence off switch, and Amazon v. Perplexity reportedly at the 9th Circuit.
In review, not yet published — These drafts have not passed our source review, so we report no figures or quotes from them yet: Hacktron's OpenAI forum-to-account-takeover write-up, a browser-extension takeover of AI agents in Comet, Edge, Opera Neon and Claude in Chrome, Dario Amodei's essay on pacing the frontier, Check Point's ChatGPT Gmail-connector finding, Perplexity's Portable Computer on RTX PCs, Cowork in Claude with Claude Docs and Claude Slides, Claude Code Projects as cloud threads, OpenAI's misalignment reporting framework, Anthropic's measurement of Claude in its own R&D, and OpenAI's Sponsored Agents test. Each is written up when it publishes.
Écarté
- A second post on DeepSeek V4's weekday-only peak window. It duplicated the September 9 post and stated V4 Flash rates that the September 10 retirement had already overtaken. Retired at review on September 14.
- A second GLM-5.3 pricing post, covering the same story as the one that published. Kept in draft at review on September 14 so the verdict update would not be written twice.
- OpenAI customer stories and corporate posts: Perplexity on GPT-6 Astra, Cognition testing Devin with Astra, scaling storage to 1B ChatGPT users, Cooley, Fyxer, AARP workshops, usage analytics and an economic research essay. No product, price or model change.
- Funding, deals and corporate finance: Superhuman buying Fathom, Mecka AI, Moonshot AI's revenue target, Temporal's $550M, Crusoe's $3.9B, Instinct's $10B target, Emerald AI, OpenAI buying Glass Imaging, Altman calling a 2026 IPO "ill-advised", and Anthropic IPO and compute reports. None moves a tool, model or price.
- Political statements and opinion on AI pacing and safety: Obama, Garry Tan, David Sacks, Speaker Johnson, The Register on regulatory capture, the BBC on insider warnings, a Suleyman podcast, a Pew survey and follow-up opinion pieces on the Amodei essay.
- Re-reports of Anthropic's September threat report, which was already covered: Chinese labs distilling Claude, Claude used to automate exploitation, Russian state hackers, and Houthi missile-guidance software.
- General security news with no AI-tool action for readers: WordPress, GitLab, JFrog Artifactory, PaperCut, Marimo, Vite, the PhantomRaven npm stealer and a weekly roundup.
- Fable 5.1 solving a 370-year-old cipher, and Astra and Fable still gaming simple variants of 2025 alignment evals. A capability demo and a research result on existing models, with no access, price or pick change.
- Baseten adding DeepSeek V4.1 Flash, and a CryptoBriefing item on the ChatGPT Pro pause. Third-party hosting of a model already in the directory, and a story we already cover.
- Hardware and off-beat items: Fujitsu's Monaka CPU, a Huawei chip, Snap Specs, Pinterest Restyle, Apple servers and more from the TechCrunch and Verge feeds.
- Community projects below the bar: the Byteshape Qwen 3.8 27B quantisation (Bonsai 2 is the vendor-backed one), OpenJev, a Qwen3-TTS fine-tune, Nvidia CUDA for Rust, Bend, and a Cloudflare security-audit skill.
- The long tail of Hacker News stories on both scan days: Show HN launches, opinion posts and third-party comparison blogs.
Our take: The verdict on this week: no directory rating moved on a published post, but three published prices and one model id no longer mean what they did. Claude Sonnet 5 stays recommended at $2/$10, now the standard rate. GLM-5.3 stays conditional, with its Flash tier at $0.15/$0.50. Anything still calling deepseek-v4-flash is running V4.1-Flash. The bigger stories are still in review and count only toward drafts: Claude Code weekly limits, security write-ups on DeepSeek Harness, Docker Sandboxes and ZCode, and launches including Gemini 3.8 Live and Qwen3.8-Omni-Flash. None has passed review yet. The directory grew by six entries: SWE-2, LiteLLM and DeepSeek Harness on September 14, then Perplexity Computer, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on September 18. Counting basis: scanned (125) sums every triage bucket from the September 14 and 18 scans (80 and 45), of which 85 were noise. The scans produced 18 news drafts. Mattered (4) counts the posts that passed review and published this week, two on September 14 and two on September 18. Verdicts changed (0): both published verdict payloads held their rating. Two other verdict writes held their rating: Cursor stayed conditional on September 14, and our Claude Sonnet 4.6 entry stayed conditional on September 18, when its summary was corrected to say Sonnet 5 costs less. A move of DeepSeek V4 Flash from recommended to conditional was staged on September 11 and never published: the retired model's entry stays recommended as its record, and its replacement, DeepSeek V4.1 Flash, is rated conditional. Coverage note: scans ran only on September 14 and 18. Week 37 was assembled on September 11 and not re-assembled, so radar items dated September 8 to 13 appear here for the first time.
— Neomanex Utility editorial
Recevez le numéro 25
Les changements de verdict et les échéances de la semaine prochaine, dans votre boîte mail.