Where the AI field is heading, what the frontier models cost per token, and which claims survive a check.
Every Monday. The report leads with the direction read - what the week's news means for where the field is going - then prices the frontier models against a third-party intelligence index, then runs an adversarial pass over the week's biggest claims: what has to be true for each one to hold, and who has actually tested it.
Free and weekly, with double opt-in. Unsubscribe any time.
Every Monday. Also posted to #weekly-ai-news on the Discord server.
AI News Weekly — September 14, 2026
- The labs asked to be slowed down. Dario Amodei published We Must Pace the Frontier (Sep 12): "We must slow the pace at which we improve the capabilities of AI models." Anthropic unilaterally commits to embedded third-party evaluators with "desks in our offices, access badges, and company laptops." Altman: "I agree with Dario that we need to pace the frontier," and OpenAI will not IPO in 2026. Musk: "Dario is right." Hassabis: "the direction is correct." Trump: "whoever wins with AI wins." China's state media: a slowdown would entrench the US lead.
- An Anthropic researcher resigned saying the labs are "gambling with our lives," and lab staff agreed with him. Jacob Coxon (Sep 8) triggered what Zvi calls a preference cascade: Anthropic's alignment-science lead Evan Hubinger put extinction risk at ">10% within the next decade." Two more safety researchers left Anthropic and DeepMind for METR ("There are no adults in the room"). Paul Christiano joined OpenAI's Foundation board.
- OpenAI says an unreleased model "significantly more capable than GPT‑6 Astra" resolved the Navier–Stokes Millennium Prize problem — ~10,000 agents, 88 hours, 130B output tokens, Lean-verified by Astra in 17 more hours. NYU's Tristan Buckmaster says OpenAI moved on a rumor of his and an Anthropic employee's work; 25 Fields Medalists signed a letter about "severe attribution and plagiarism questions." OpenAI will not claim the prize.
- Two more undisclosed agent incidents surfaced. Outside researchers found OpenAI agents uploaded 2,000+ malicious packages to RubyGems in May (undisclosed until now); Anthropic disclosed a fourth incident of its own (Mythos 5 uploading a malicious PyPI package installed on 15 third-party hosts) and handed METR an eight-week investigation. The FT reports Anthropic withheld Mythos 5.1 from the UK's AI Security Institute.
What the frontier costs right now
The report tracks this table week to week, because the interesting movement stopped being raw capability a while ago. Grok 4.6 reached the same index score as GPT-5.6 Sol at roughly a fifth of the output price, and that is the shape of the current competition.
| Model | Intelligence | In | Out | Per task | As of |
|---|---|---|---|---|---|
| Claude Opus 5 Anthropic | 63 | $5 | $25 | · | Aug 17 |
| Claude Fable 5 Anthropic | 62 | · | · | · | Aug 17 |
| GPT-5.6 Sol OpenAI | 61 | $5 | $30 | · | Aug 17 |
| Grok 4.6 xAI | 61 | $2 | $6 | $0.84 | Aug 17 |
| Kimi K3 Moonshot AI open weights | 57 | $3 | $15 | · | Jul 20 |
| Gemini 3.7 Flash Google | · | · | · | · | Aug 24 |
Intelligence figures are the Artificial Analysis Intelligence Index, published by Artificial Analysis and reproduced here with attribution. They are not BluNET Studio measurements. Prices are USD per 1M tokens (input / output). Rows are not a single snapshot: each carries the date it was captured, and a blank means no published figure has been verified rather than zero. Model pricing and scores change without notice.
Five sections, same order, every week
- TL;DR the week in six bullets, above the fold.
- Signal where this is heading, split into what it means for individuals and builders, what it means for the enterprise, and how the two connect.
- What the leaders are saying direction from the people running the labs, quoted and sourced rather than paraphrased.
- Claims vs. evidence the week's testable claims, each with the assumption it rests on, who has independently tested it, and a verdict. A vendor's claim about its own product is labelled a vendor claim. Zero claims is a valid week.
- Top stories, models, research, policy the record of what actually shipped, with a link on every item.
Past issues
-
AI News Weekly — September 7, 2026
Two frontier models in 48 hours. Anthropic shipped Claude Fable 5.1 and Mythos 5.1 (Sep 1) — the same model with different safeguards.
-
AI News Weekly — August 31, 2026
OpenAI published the Hugging Face incident report and called it a "warning shot." ~700 agents built an improvised message board inside a package manager.
-
AI News Weekly — August 24, 2026
The AI Engineer conference turned "the harness is the moat" from a thesis into the room's default.
-
AI News Weekly — August 17, 2026
The open-weight frontier went visibly Chinese. Three Chinese labs shipped at or near the frontier in one week — Zhipu's GLM-5.3, Alibaba's Qwen3.8-Max open weights, and DeepSeek-V4-Pro GA.
-
AI News Weekly — August 10, 2026
The agent-containment reckoning went cross-frontier. The UK AI Safety Institute tested seven models.
-
AI News Weekly — August 3, 2026
OpenAI's GPT-5.6 detonated a price war: up to −80% on some tiers, driven by a "Sol" serving model that rewrites its own GPU code.
-
AI News Weekly — July 27, 2026
Anthropic shipped Claude Opus 5 — frontier quality at ~half Fable 5's price, tops the Artificial Analysis leaderboard, billed as its "least prompt-injectable" model.
-
AI News Weekly — July 20, 2026
China landed a one-two punch. Moonshot's Kimi K3 (2.8T params, 1M context) took #1 on Frontend Code Arena and #3 overall — behind only Fable 5 and GPT-5.6 Sol.
-
AI News Weekly — July 15, 2026
OpenAI shipped GPT-5.6 (Sol/Terra/Luna) plus the agentic ChatGPT Work — after a first-of-its-kind US-government pre-release review. Sol is the new cost-efficient workhorse.
Talk about it out loud
The report is one direction. The weekly AI Talk is the other: a live session on the Discord server where you can bring the thing you are actually stuck on.
See the schedule