News section
Foundation and frontier model releases, updates, and capability changes.
14 Important · 5 verdict changes · 10 FYI · 1 watching · 30 total
ImportantModels ·
SpaceXAI says Grok Voice Transcribe 2.0 will soon become the default in its Speech-to-Text API and that 1.0 will be deprecated "in the coming weeks", with no date for either. Pin grok-voice-transcribe-1.0 and test 2.0 on your own audio before your transcripts change under you.
Read the full callImportantModels ·
Google began rolling out Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking to the Gemini API and Google AI Studio on 15 September 2026. Google says Extended Thinking ranks first on Artificial Analysis's Speech-to-Speech Quality Index at 82.6, and it is the one we recommend you trial first.
ImportantModels ·
PrismML's Ternary Bonsai 2 27B stores Qwen3.8 27B at 1.76 effective bits per weight: 5.9GB, more than 9x smaller than full precision, at a claimed 98.2% of the full model's aggregate score. The weights are out under Apache 2.0, and every figure is PrismML's own.
ImportantModels ·
Anthropic released Claude Sonnet 5.5 on 28 September at $2 input and $10 output per 1M tokens, the same as Sonnet 5. Cache reads launched at $0.20 and now list at $0.10. Claude Code 2.1.284 made it the default Sonnet on the Anthropic API. If you run Sonnet with thinking off, switch to the new between_tools setting before you migrate. Every benchmark is Anthropic's own.
ImportantModels ·
Since 17 September, verified life-sciences teams on the Claude API, Enterprise or Team plans can apply to Anthropic's Life Sciences Verification Program to relax the biology safeguards that block legitimate research. Standard Use launched on Mythos 5.1, Opus 5 and Sonnet 5 and now lists Opus 5.5; High-risk Use covers Opus 5 and Sonnet 5, while Mythos stays limited to a small vetted set.
ImportantModels ·
Alibaba's Qwen3.8-Omni-Flash takes text, image, audio and video in, has a 1M-token context, and is callable on Model Studio as qwen3.8-omni-flash. Qwen claims a price cut of more than 98% on audio input, but Qwen's launch post shows its rates only as an image and Model Studio's documentation lists no price, so check your console before moving a workload.
ImportantModels ·
Google shipped gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts to the Gemini API and AI Studio on 23 September at $9.00 and $6.00 per 1M audio output tokens, less than half of Gemini 3.1 Flash TTS Preview's $20.00. Every 3.8 rate doubles on 1 January 2027, so we rate both conditional.
ImportantModels ·
OpenAI made GPT-Live-1, its full-duplex voice model, generally available in the API on 10 September at $0.05 per minute for the voice session, with telephony support. The reasoning runs on a backend model you pick and pay for separately, and OpenAI publishes no all-in per-minute cost. GPT-Live-1 stays conditional: the no-API caveat is gone, and cost is the new one.
ImportantModels ·
DeepSeek released V4.1-Flash on 10 September, a 552B MoE with open weights, 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. Same id, different model: re-run your evals.
GGImportantModels ·
Google's third Flash release in six weeks lands at the same $0.75/$3.75 intro price as 3.7 Flash, with a caveat: 3.8 'works harder' and may burn more tokens. The gated 3.8 Flash Cyber goes to trusted defenders under the new Fairwind Program.
GImportantModels ·
Gemini 3.7 Flash lands three weeks after 3.6: better on every benchmark Google published, at half the intro price through 2026. The upgrade path is obvious.
M
ImportantModels ·
MAI-Code-1.1-Flash is in Copilot: native vision, 22% better Terminal-Bench, a quarter of the cost. MAI-Code-1-Flash deprecates Sep 10.
GImportantModels ·
GLM-5.3 posts a more-than-6× Terminal-Bench 3.0 jump and leads CyberGym at 84.5%, ahead of Mythos 5 and GPT-5.6 Sol. Open weights land in two weeks, not today.
ImportantModels ·
The week of August 10 is here. Alibaba confirmed on Aug 3 that open weights for Qwen3.8-Max (2.4T MoE) and a 27B dense companion would drop "next week." As of Monday Aug 10, no HuggingFace repository exists, the license is undisclosed, and the only delivery date comes from a secondary source. Here's what's confirmed, what's pending, and why the 27B matters more than the flagship.
Verdict changedModels ·
Our directory entry for Claude Fable 5 said credit-only, no subscription path, eight weeks after we reported the opposite. Anthropic's plan table puts Fable at 50% of weekly limits on Max 5x, Max 20x and Team premium seats. Here is the correction, and what it does not change.
Verdict changedModels ·
Claude Sonnet 5's $2/$10 per 1M token launch price was due to rise to $3/$15 on 1 September 2026. Anthropic cancelled the increase. Sonnet 5 stays recommended, now a third cheaper than Sonnet 4.6 and 40 percent of the Opus rate, with one tokenizer caveat.
Verdict changedModels ·
GPT-6 Astra is real, expensive, and genuinely ahead on computer use, terminal work, cyber and long context on OpenAI's own numbers. It is not a general-purpose upgrade, and there is no vendor-neutral general-intelligence score published for it at all. Route to it by task.
Verdict changedModels ·
Qwen3.8-27B's open weights are live under Apache 2.0: a 27B single-GPU model with 262K context and vendor-claimed coding gains. Pending → conditional.
Verdict changedModels ·
Grok 4.6 is live at $2/$6, scoring 61 on the AA Intelligence Index, tying GPT-5.6 Sol for third, behind Claude Opus 5 and Claude Fable 5. Verdict: pending → conditional.
FYIModels ·
Amazon Bedrock added xAI's Grok 4.7 on 28 September 2026 at $2.00 input and $6.00 output per 1M tokens on Global cross-Region inference ($2.20/$6.60 on US Geo), and its Flex tier bills at half that. The catches: no in-Region endpoint, no server-side tool use, and Flex does not work through the Converse API.
FYIModels ·
DeepSeek's rate card now excludes Chinese public holidays from its weekday peak window, so those days bill at half the peak rate. Per-token rates are unchanged, and the V4 Pro routing question is still open.
MFYIModels ·
Inception released Mercury 2.5, a diffusion language model, at $0.20 input and $0.75 output per 1M tokens, discounted 80% to $0.04 and $0.15 at launch with no end date. Inception claims a 260K-token context, 1,107 tokens per second and a 40% intelligence gain over Mercury 2. Every figure is Inception's own.
FYIModels ·
Google added Live Avatar to Gemini 3.8 Live on 24 September: a near real-time talking avatar with lip-sync in 97 languages, available in Gemini Enterprise. Custom avatars from your own reference images need enterprise allowlisting, audio and video carry a SynthID watermark, and Google published no price.
FYIModels ·
DeepSeek's rate card scopes its 2x peak window to Monday through Friday. Our entries said "seven daily peak hours", which overstated the expensive window by 14 hours a week on both V4 Pro and V4 Flash.
KFYIModels ·
The Institute of Foundation Models shipped a six-model Apache-2.0 agentic fleet on September 3 (from a 0.9B edge model to a 375B-A23B MoE), with the training lifecycle opened end to end. All claims are IFM-reported until artifacts are independently verified.
FYIModels ·
Qwen3.8-Max-0902 is a hosted-API checkpoint refresh of the Qwen3.8-Max flagship: $2/$6 pricing and 1M context, still not the promised open-weight drop.
FYIModels ·
OpenAI now formally classifies Astra as its first 'Critical'-threshold cybersecurity model: ExploitBench 100%, two zero-days found during evaluation. Launch is 'soon,' with advanced cyber access gated to alpha testers before Daybreak Blue.
FYIModels ·
Qwen3.8-Max publishes $2/$6 per-token pricing on QwenCloud, replacing the credit-only Token Plan. Open weights promised this week, but not yet shipped.
FYIModels ·
Meta released Muse Glimmer, a 30B open-weight model under Apache 2.0, designed for always-on local agents on a single consumer GPU. 131K+ context. The first open-weight model purpose-built for local agentic autonomy.
WatchingModels ·
Grok 4.6 missed its August 7 launch. Musk's Aug 4 call pushed the window to Aug 10–14. No model card, benchmarks, or API. The model stays pending.
Never need to catch up again
The weekly delta: only verdict changes and act-now items. No digest filler.