Quick answer

Claude Sonnet 5.5 (Sept 28), GPT-6.1 Sol (Sept 29), and Gemini 4 Argon (Sept 30) all list $2 per million input tokens and $10 per million output. Beyond the price they diverge: Sonnet 5.5 and Sol are generally available now, Argon is gated to the Fairwind Program; Sonnet and Argon offer 1M-token context, Argon alone a 1M-token output limit; Sol's cached input is $0.10, Sonnet's cache reads $0.20, Argon's cached input $0.10; Argon's price doubles after its promotion, the other two are standard rates. All three vendors claim near-flagship performance; none of it is independently verified yet.

For the first time the three labs have landed on the same price for their workhorse model in the same week. That makes the comparison unusually clean: the question is no longer what you pay but what you get, and when.

At a glance

  • Sonnet 5.5 (Anthropic): $2/$10, cache reads $0.20 with a 512-token minimum, batch $1/$5; 1M context, 128K output; available on the Claude Platform, Bedrock, Vertex, and Azure; >30% faster than Sonnet 5 and up to 30% fewer tokens per task, per Anthropic
  • GPT-6.1 Sol (OpenAI): $2/$10, cached input $0.10; available in the API, Codex, and ChatGPT Work (not the standard chat picker); near-Astra on coding, computer use, and professional tasks at one-fifth the price, per OpenAI; Ultrafast mode coming
  • Gemini 4 Argon (Google): introductory $2/$10 rising to $4/$20, cached input $0.10; 1M context and 1M output; text, image, video, speech in; Fairwind Program only at launch, paid API and AI Ultra next with no date

Availability decides a lot

Only two of the three can be used by a general developer this week. Sonnet 5.5 is live across every major cloud. Sol is live in OpenAI's API and Codex. Argon is for vetted cyber defenders, with the public API returning not-found for the model ID on October 1. If your decision is for this quarter, it is between Sonnet and Sol, with Argon as the one to re-evaluate when Google opens it.

Where each is strongest, by the vendors' own framing

  • Sonnet 5.5: agentic coding, with a reported 70.6% on Terminal-Bench 4.0 that Anthropic says edges Opus 5.5; the model most Claude Code and coding-agent users will run
  • Sol: the default across OpenAI's surfaces, with the ecosystem (Codex, plugins, Dots, ChatGPT Work) as the real advantage
  • Argon: very long runs, with the 1M-token output and Long Decode Continuation, and Google-reported leads on legal, finance, and software-engineering agent benchmarks

API behaviour to check before switching

  • Sonnet 5.5 returns 400 errors for explicit thinking budgets, sampling parameters, prefill, forced tool choice, and disabled thinking; thinking is always on
  • Sol is accessed through OpenAI's current API shapes; existing GPT-6 code should move with a model-ID change, but Work-only availability in ChatGPT limits consumer testing
  • Argon's API is not yet public; plan against Google's published pricing and the 1M output limit, not against hands-on behaviour
  • Caching rules differ: Anthropic's 512-token minimum versus OpenAI's and Google's automatic cached-input discounts; the cheapest model for your workload depends on how much of your prompt repeats

A routing suggestion

Run coding agents on Sonnet 5.5. Run ChatGPT Work, Codex, and anything that leans on OpenAI plugins on Sol. Keep Argon on the roadmap for hours-long agent runs on Google Cloud. Put a gateway in front so the routing is a configuration change, and re-test all three on your own tasks once independent benchmarks appear.

Related reading

Claude Sonnet 5.5 released Gemini 4 Argon announced LLM gateways compared

Bottom line

Same price, different availability, different strengths. Sonnet 5.5 for coding today, Sol for the OpenAI ecosystem today, Argon for long-horizon work when you can get it. The convergence on $2/$10 is the real news: frontier-class capability at that price is now the floor, not the ceiling.