AI Got Cheaper, Bigger and Harder to Trust This Week

Seven days of AI news came down to six numbers this week. We put them in a two-minute film, and every figure in it traces back to a source you can open. You can watch it on this page, or as a YouTube Short.

Cheaper. Anthropic released Claude Haiku 5.5 on October 7 at $0.10 per million input tokens for prompts up to 100,000 tokens. The model it replaces, Haiku 4.5, cost $1.00. That is a tenth of the old price on the prompts most people send: Anthropic says about 90% of requests to the last Haiku were under that size. Above it the price is $0.50, and the company's own blended figure is around 75% cheaper to run.

A day later OpenAI moved everyone on the Free and Go plans of ChatGPT to GPT-6 Luna. Answers can now include charts, forms and tappable buttons, which OpenAI calls Intelligent UI. Paid plans get a different model, GPT-6 Sol.

Then there is the gap. Bloomberg reported on October 4 that top Chinese models trail their US rivals by 3% on benchmark scores, after DeepSeek's V4.1 Flash scored 81.1 on LiveBench against 83.4 for Anthropic's leading model. The same Bloomberg Intelligence analyst put that gap at about 9% in May. It is one snapshot of one benchmark, and the film says so on screen.

Bigger. Mistral previewed Large 4, which it calls le Chonk, on October 6. It has 1 trillion parameters. How many of them are active for any one answer depends on which Mistral source you read: the launch post said 49 billion and the company's page now says 52 billion. The film says about fifty. Mistral says the weights will be released by the end of the month.

Three numbers to act on

Harder to trust

Tavus ran its own test of a new video model called Griffin. People were told they would be matched with another participant for a one-minute video call. Of the 54 who spoke with it, 26 believed their partner was a real person, which is 48%. Tavus announced that on October 1, a few days before the week this film covers. We kept it in because Koda's Deep Dive on it ran on October 10. The study is small, and the company graded its own work.

SemiAnalysis counted 857 model releases from nine leading Chinese developers between 2021 and September 15, 2026, then looked for a published safety result that could be matched to each one. It found 31, or 3.6%. For 813 of them it found no safety disclosure at all. The report gives no matching figure for US labs, so the film makes no comparison.

Why a finance team should care

I run AI on real finance work every day, and I read this week as two instructions.

First, price is a live variable again. If a model at a tenth of the cost writes your variance commentary as well as the one you pay for, the saving is real. Only your own test will tell you, because a benchmark is not your workload.

Second, the cost of producing something convincing keeps falling faster than the cost of checking it. A call, a report or a model card can look complete and have nothing behind it. Decide now who checks, and against what.

Sources. Anthropic, Introducing Claude Haiku 5.5, October 7, 2026. OpenAI, GPT-6 and Intelligent UI in ChatGPT, October 7, 2026. Bloomberg, US Lead in AI Over China Narrows After DeepSeek Gains, October 4, 2026. Mistral, Introducing Mistral Large 4, October 6, 2026. Tavus, Griffin, October 1, 2026. SemiAnalysis, Beijing Will Not Pace the Frontier, October 8, 2026. We checked all six live on October 11. The Earth shots in the film are illustrations, not satellite footage.