OpenAI and Anthropic now ship a new model about every 11 days

Ten days brought more than ten launches

Between Sept 29 and Oct 8, releases kept coming from big labs and open-weight teams alike. In order of posting:

• OpenAI's sol 6.1 arrived with cheaper cached tokens, which @kunchenguid called a big deal for the cost of long-running work (Sept 29).
• Gemini 4 Argon (High) took #1 in Text Arena with 1,525 points, 20 ahead of Claude Opus 4.6 (High), at a blended $8 per million tokens, per @arena (Sept 30).
• Grok 4.7 reached Amazon Bedrock and Google's Gemini Enterprise Agent Platform, at $2 input and $6 output per million tokens on the xAI API, per @tetsuoai (Oct 4).
• Mistral Large 4 opened as a preview with 1T parameters and 49B active; @GuillaumeLample says the final version and weights come before month's end (Oct 6).
• The same 24 hours brought Aleph Alpha's Kolibri-1, ElevenLabs v4 Turbo and BFL's FLUX 3, as @SebJohnsonUK listed (Oct 6).
• GPT-6 with Intelligent UI began rolling out in ChatGPT, where changing a dinner guest count updates the shopping quantities, per @kimmonismus (Oct 7).
• Nano Banana 2.1, Google's new image model, ranks #4 on both AA-Image-T2I v2.0 and AA-Image-Editing v2.0, with its largest gains in Retail & Ecommerce, and it already sits inside Google Ads, per @ArtificialAnlys (Oct 8).
• Google's Gemini agent debuted before 700+ Google Cloud customers as one prompt box that answers questions, makes images, and writes and runs code, per @sundarpichai (Oct 8).
• Underdog Saluki 27B, built on Qwen 3.8 27B, keeps 96% of its benchmark performance at almost 7x smaller, under Apache 2.0, per @UnderdogAI (Oct 8).

More is queued. @DrPengShen wrote on Oct 5 that DeepSeek V5, GLM 5.5, Kimi k4 and Qwen4 are on the way.

The leader changes within days

Three pairs show how short the shelf life is. Sol 6.1 shipped only a few days after sol 6, enough for @kunchenguid to feel sol 6 had gone out early. Nano Banana 2.1, released Oct 6, replaced Nano Banana 2 on the Gemini API at half its price: $0.0336 per image at 1K. Google's previous release, Gemini 3.8 Flash (High), sat at #11 in Text Arena before Argon took #1, and Google's spot in Code Arena: WebDev went from #29 to #8.

A top spot this week is a snapshot.

Rankings also leave things out. The same day Argon topped the board, @firstadopter shared a Bloomberg report, citing people with direct access, about doubts inside Google before launch. Gemini 4 did well on industry benchmarks, but "it does less well when employees actually put it to work," the report said, with trouble on certain coding tasks.

That makes rank the wrong number to track. What matters for a store is whether a new version does one of its jobs better, such as product images, than the model it already uses.