AI Model Roundup: Sep 7-13, 2026
The week's biggest claim had nothing to do with chatbots: OpenAI said an internal agent system produced a machine-checked proof on the Navier-Stokes Millennium Prize problem. On the product side, GPT-6 Astra completed its commercial rollout, DeepSeek V4.1 Flash launched with a price cut, Mercury 2.5 reached general availability, and Artificial Analysis rescaled its Intelligence Index. Benchmark figures below are vendor- or third-party-reported.
The Week at a Glance
- Model launch: Mercury 2.5 (Inception) GA on Sep 8, diffusion-based LLM, 1,107 tok/s, 260K context, $0.20 / $0.75 per MTok.
- Open weights: DeepSeek V4.1 Flash launched Sep 10 with native vision and lower prices.
- Full rollout: GPT-6 Astra generally available on ChatGPT tiers, API, Azure, and Bedrock. Enterprise plugins added.
- Benchmarks: Artificial Analysis Intelligence Index v4.3 published Sep 7. Top scores: Claude Fable 5.1 and GPT-6 Astra tie at 53.
- Science claims: OpenAI said an internal agent system solved the Navier-Stokes Millennium Prize problem (Sep 8). Google DeepMind released AlphaGenome Atlas.
Model Releases and Announcements
Mercury 2.5 GA (Inception Labs, Sep 8)
Inception launched the diffusion-based Mercury 2.5 as generally available on September 8. Claimed specs: 1,107 tok/s on NVIDIA GPUs, 260K context window, tunable reasoning. List price is $0.20 per million input tokens and $0.75 per million output tokens. Inception applied an 80% launch discount bringing it to $0.04 / $0.15.
Caveat: The Mercury 2.5 Preview discount expired September 8 at 07:00 UTC, reverting that endpoint to list price. Inception has not published independent benchmarks for 2.5. (Source: Inception blog.)
GPT-6 Astra Full Commercial Rollout (OpenAI)
GPT-6 Astra became generally available across ChatGPT Plus, Pro, Business, and Enterprise, plus the API (model name gpt-6-astra, $10 / $50 per MTok, ~$1 cached input, 2x for fast mode), Azure, and Bedrock. The rollout had started as a phased debut on September 3, but OpenAI's launch page is timestamped September 10. On September 9, Astra became the default for ChatGPT Work and Codex (admin opt-in). Enterprise plugins added include Oracle Analytics, Power BI, Navan, and Avalara. On September 11, OpenAI Developers advised reworking AGENTS.md and skills for Astra. (Sources: OpenAI launch page, OpenAI Developers.)
DeepSeek V4.1 Flash (Sep 10)
DeepSeek released V4.1 Flash, a 552B-parameter MoE with a new encoder-decoder design (8B active parameters for input, 16B for output), native image understanding, and open weights on Hugging Face. Off-peak pricing is $0.15 per million input tokens (cache miss) and $0.60 per million output, with peak hours at double (pricing). DeepSeek initially said all V4 Pro API traffic would route to V4.1 Flash from September 14; see next week's roundup for how that turned out.
Benchmark Developments
Artificial Analysis Intelligence Index v4.3 (Sep 7)
The new index version ranks models on a composite score. Top-level results (max configuration with fallback where applicable):
| Model | Score | Cost per task |
|---|---|---|
| Claude Fable 5.1 (max w/ fallback) / GPT-6 Astra (max) | 53 (tie) | $7.63 / $3.26 |
| Claude Opus 5 | 51 | n/a |
| Claude Fable 5 | 50 | n/a |
| Muse Spark 1.3 | 48 | n/a |
| GPT-5.6 Sol | 47 | n/a |
Open-weight leaders: GLM 5.3 Flash (42), Qwen3.8 2.4T A95B (40), DeepSeek V4 Pro 0813 (36). Astra (max) costs $3.26 per task versus $7.63 for Fable 5.1 at equal score. (Source: AA article.)
Pricing, API Changes, and Provider News
- Mercury 2.5 pricing shift: Preview discount ended Sep 8; GA launch discount (80%) began. See above.
- OpenAI Navier-Stokes claim (Sep 8): OpenAI says an internal multi-agent system (about 10,000 concurrent agents running for roughly 88 hours, plus 17 hours of formalization in Lean) produced a proof of finite-time blowup for the 3D Navier-Stokes equations with a smooth external force, which corresponds to statements C and D of the official Millennium Prize formulation. The proof and the Lean formalization are published, so the core argument can be machine-checked. OpenAI says it will not claim the $1M Clay prize; mathematicians are still reviewing the result. (Sources: OpenAI, Quanta Magazine.)
- Google DeepMind AlphaGenome Atlas (Sep 8): Precomputed 1-petabyte catalogue of predicted effects of ~9 billion single-nucleotide variants; a new AVI score; free web portal for noncommercial use, plus API and Antigravity skill; commercial access via Google Cloud to follow; accompanied by a Nature paper. (Sources: Google blog, SiliconANGLE.)
What to Watch Next Week
- Whether DeepSeek goes ahead with routing V4 Pro traffic to V4.1 Flash on September 14.
- Continued adoption of GPT-6 Astra across enterprise tiers.
- Expert review of the Navier-Stokes proof and whether the Clay Institute changes the problem's status.
Takeaway
At the top of the index, cost per task now separates models more than scores do: Fable 5.1 and GPT-6 Astra tie at 53, but Astra costs $3.26 per task against $7.63. If you run frontier-level agents at volume, that gap is worth an eval. For cheaper work, V4.1 Flash and Mercury 2.5 both moved the price floor this week.
For earlier launches, see AI Model Roundup: Aug 17-23, 2026. For model rankings across categories, visit our Overall models page.