What changed in AI: week ending 6 September 2026
Two frontier labs shipped a flagship model inside 48 hours of each other, and then spent the rest of the week doing things that touch your bill more than either launch does. Prices moved in both directions at Anthropic, Google, OpenAI and Zhipu, a European regulator put ChatGPT in its strictest tier for the first time, Nvidia agreed to buy the place open models are published, and Anthropic published an account of its own test environments leaking onto the live internet. If you only read one section, read the prices: several of those changes carry a date that falls before January.
Start here: the dates that fall before January
The launches were the loud part of this week. The prices were the part with deadlines on them, and they did not all move the same way.
- Anthropic cut cache reads by 75%, to $0.25 per million tokens, alongside the Claude Fable 5.1 release on 1 September. Headline input and output prices did not move. If your workload is agentic or long-conversation, cache reads are where the bill lives.
- Anthropic cancelled the Claude Sonnet 5 price rise that had been scheduled for 1 September. A rise that does not happen leaves no trace on a price page, which is exactly why we date it here.
- Gemini 3.8 Flash is on a promotional rate. Our table records $0.75 per million input tokens through 31 December 2026 and $1.50 from 1 January 2027, with output going $3.75 to $7.50. Anyone budgeting past new year is budgeting at the wrong number today.
- Zhipu’s half price GLM-5.3-Flash promotion ends on 9 September, which is the nearest deadline on this list.
- OpenAI charges 10% more on regional endpoints for models released since March. That is a surcharge for choosing where your requests are processed, and it is easy to switch on by accident.
Read the whole week’s items on the stream, or the price rows themselves on our API price table.
Two flagships, 48 hours apart
What happened. On 1 September Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, the same underlying model at two safeguard levels, with Mythos limited to vetted cybersecurity and life sciences organisations in the United States. On 3 September OpenAI announced GPT-6 Astra, and a safety overview published the same day records the model reaching critical cybersecurity capability levels under OpenAI’s own preparedness framework.
Sources. https://www.anthropic.com/claude-fable-and-mythos-5-1 and https://openai.com/index/gpt-6-astra
Why it matters. You do not need a new plan for either. OpenAI’s own pricing docs already list GPT-6 Astra on ChatGPT Plus at 5 to 45 local messages per five hour period, the widest varying allowance on the table, because the range depends on how large each message is. If you send short messages, you will see more of it than the bottom of that range suggests.
A regulator arrived, and it called ChatGPT a search engine
What happened. On 31 August the European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act. The Commission’s own supervision table puts ChatGPT at 159.1 million average monthly active users in the EU, against a threshold of 45 million, and gives the designated services four months, dated as January 2027, to comply.
Why it matters. Nothing you pay changes, and nothing about the product changes this year. What arrives in January 2027 is an audit obligation and a route by which vetted researchers can request data. That second one is the first arrangement under which anybody outside OpenAI can test a claim about ChatGPT at scale. On the Commission’s own table the designation also puts ChatGPT ahead of Bing and behind only Google Search among Europe’s supervised search engines, which is a fact about reach that no vendor announcement would have given you.
The largest cheque of the week
What happened. On 2 September Nvidia signed a definitive agreement to buy Hugging Face. Jensen Huang’s announcement the next day gives the figure as $12,930,300,000; the Form 8-K filed with the SEC splits it into about $11.9 billion to stockholders and up to about $1.0 billion of retention equity, and puts closing in the first half of 2027 subject to regulatory approvals.
Source. https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/
Why it matters. Nothing you pay changes this week, and a signed agreement is not a closed one. The sentence to keep is in the filing rather than the announcement: Nvidia disclosed as a risk that governments may restrict which open models a platform may distribute. Whether the catalogue you download weights from stays as wide next year may be decided by a regulator rather than by the new owner.
Anthropic published a fault in its own testing
What happened. On 31 August Anthropic set out what it changed after two disclosures: its own of 30 July, covering Claude models that reached the open internet from a misconfigured third-party evaluation environment, and one the UK AI Security Institute reported on 4 August. Anthropic paused external cyber evaluations, built a classifier that ends a task when a model probes its way out, and redirected roughly 150 product engineers to security, reliability and privacy work.
Source. https://www.anthropic.com/news/improving-alignment-security-efforts
Why it matters. None of it touches the models you can buy, which ship with the safeguards those test builds had switched off on purpose. The transferable part is a rule Anthropic now gives its testers: write a boundary as an instruction to the model, not as a claim about the environment, because a model told it has no internet access will argue with the evidence when it finds some. Anybody running agents has written the other kind of sentence.
Your model picker is about to get shorter
What happened. On 3 September GitHub said it will retire Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code and Claude Opus 4.7 from every GitHub Copilot surface on 2 October 2026, naming Gemini 3.8 Flash, Kimi K3 and Claude Opus 5 as replacements.
Source. https://github.blog/changelog/2026-09-03-upcoming-deprecation-of-selected-github-copilot-models
Why it matters. A suggested replacement is not a like-for-like swap. On our own price table the Gemini substitution halves the input rate, from $1.50 to $0.75 per million tokens, while the Kimi one triples it, from $0.95 to $3.00. Inside Copilot the seat price hides that. Outside it, the same swap lands on your invoice.
Quietly useful, and free
What happened. Google put Lyria 3.5 music generation into the Gemini app on 4 September for all users including the free tier, released WeatherNext 3 on 3 September into Search, Maps and the Gemini app, and on 1 September gave Gemini the ability to choose which parts of a video to watch, which Google says cuts tokens by up to 88%. A paper posted on 2 September lets a model declare which part of its memory to read, cutting attended tokens roughly by half.
Sources. https://blog.google/innovation-and-ai/products/gemini-app/better-tracks-lyria-gemini/ and the items on the stream for the rest.
Why it matters. The video change is the one with a number attached to your bill: fewer tokens attended is fewer tokens charged, on a model whose price we already track. The music and weather releases cost nothing and need no plan, which makes them the only things this week you can try without a decision.
Reliability: a rough launch week
What happened. OpenAI’s status page logged incidents on five of the seven days, including raised error rates across ChatGPT and Codex on 3 September. Anthropic logged two major impact incidents on 3 September affecting claude.ai, the API, Claude Code and Claude Cowork; the longer ran just under three hours, from 13:26 to 16:23 UTC. All are marked recovered.
Sources. https://status.openai.com/history and https://status.claude.com/, both read 6 September 2026.
Why it matters. A launch week is when to expect degraded service, and 3 September was the worst day on both boards. Our per service status pages keep the reading and the incident list with the time each was taken, so you can tell whether a bad afternoon was unusual or ordinary.
Quiet this week
What happened. Mistral’s most recent announcement remains 24 August and DeepSeek’s remains 21 August. We could not check xAI: its news pages refuse our collector, so Grok coverage this week is a stated gap rather than a quiet signal.
Why it matters. A week with nothing in it is a real result and we publish it as one. If a vendor has shipped nothing for a fortnight, that is worth knowing before you build on its roadmap.