This article was drafted together with AI (Claude) and reviewed by an editor. Anthropic is the company that makes Claude, which was used to write this article. Because it may have a stake, we label every figure as "announced by the company," and we treated news about other companies by the same standard.
Correction (2026-10-03) The first version of this article said we "could not confirm a Gemini 4 launch in Google's official materials." Google announced Gemini 4 Argon on its official blog on September 30. We corrected the title, the opening paragraph and section 3, and added the source.
This week (September 28 to October 3) there were three major announcements. From OpenAI and Anthropic, the message about lower prices came before the message about higher performance, while Google chose to open its strongest new model to a limited group first.
1. OpenAI — GPT-6.1 Sol
Introducing GPT-6.1 Sol, OpenAI describes it as "intelligence close to the previous top model (GPT-6 Astra) at a much lower cost." Developer pricing is $2 per million input tokens, $10 for output, and $0.10 for repeated input. OpenAI announced that on coding and work-automation evaluations it scored about the same as Astra at roughly one-fifth of the cost. It also says the factual error rate at low reasoning effort fell from 11.4% to 7.7%. (All figures are those announced by OpenAI.)
Worth noting: According to OpenAI's page, it opened first in ChatGPT's "Work" and in Codex, and not yet in regular chat. It is available on paid plans (Plus, Pro, Business, Enterprise, Edu).
2. Anthropic — Claude Sonnet 5.5
Anthropic describes Sonnet 5.5 as "more than 30% faster than the previous Sonnet 5, and up to 30% cheaper on most tasks." Per-token prices are unchanged ($2 input, $10 output); the explanation is that it finishes the same job with fewer tokens, so the cost goes down. On the 22nd of the same month, the higher-tier Opus 5.5 was released, and Anthropic says it "runs at 40% lower cost than Opus 5 on typical work." Anthropic positions Sonnet 5.5 for well-defined everyday tasks such as making documents, slides and spreadsheets and fixing bugs, and Opus 5.5 for complex work that requires judgment. The French IT outlet Next.ink pointed out that the unit price did not go down, and summed it up as a structure in which the price falls through efficiency.
3. Google — Gemini 4 Argon: The Strongest Model Goes to a Few First
Google announced Gemini 4 Argon on its official blog on September 30. It is first rolling out to trusted cyber defenders through Google's Fairwind Program, and Google says it will widen access, starting with paid API customers and Google AI Ultra subscribers, after refining its guardrails. Where the other two led with price, Google led with "opening the strongest model carefully." We did not find independent verification of the performance figures beyond Google's own announcement, so we do not repeat them.
Good to know: In the developer model documentation updated on October 1, the stable models you can use right now are Gemini 3.8 Flash (for long-running development work and agents), Nano Banana Pro for images (4K, text rendering) and Gemini 3.8 Live for voice. Gemini 4 Argon is also listed on Arena, the site that ranks models by human votes. For details, see the piece we posted this week, "Who Uses Which AI the Most."
Our Interpretation
Fact: Two companies put forward "the same job, cheaper" in the same week.
Interpretation: It is not because the models got better but because the cost of using them went down that the options for individuals and small companies change. For work you run every day, "the cost of handling one item" has become more important than performance scores.
What we would do: Before changing the tools you use now, pick three tasks you do every week, give them to both models in exactly the same way, and write down the time and the results. The method is in our piece "Which AI for Which Job."
Further reading
This article was written after directly opening and checking the official announcements and reports below. The figures are each company's announcements and have not been independently verified.
- openai.comIntroducing GPT-6.1 Sol (OpenAI)English
- anthropic.comClaude Sonnet 5.5 (Anthropic)English
- anthropic.comClaude Opus 5.5 (Anthropic)English
- next.inkClaude Sonnet 5.5 : plus rapide, plus efficace, mais pas moins cher (Next.ink)French
- ai.google.devGemini models (Google AI for Developers)English
- blog.googleGemini 4 Argon: our next era of frontier intelligence (Google)English