What a new model does, what it costs to run, and which of its headline numbers belong to the lab that built it. Every figure carries its source, its date and the configuration it was measured in.
Google DeepMindReleased Sep 30, 2026
Treat Gemini 4 Argon as a model to watch and not yet one to plan on: outside evaluators confirm four of Google's eight launch claims at high effort, but no customer outside a set of Fairwind security teams can get access, Google gives no date for anyone else, and any budget written today should assume the $4/$20 price that follows an introductory period of unstated length.
Rate card read Oct 1, 2026Updated Oct 1, 2026
AnthropicReleased Sep 28, 2026
Run Claude Sonnet 5.5 at high effort, its API default, where it scores level with GPT-6 Sol's best for about Sol's cost per task or less; at xhigh and max, Claude Opus 5.5 at a lower setting scores about the same or better for about half the cost per task or less on both outside evaluators that have published, so the max-effort Terminal-Bench headline describes the most expensive way to use Sonnet 5.5.
claude-sonnet-5-5Rate card read Sep 28, 2026Updated Oct 1, 2026
ComparisonAnthropic and OpenAIReleased Sep 22, 2026
Run Claude Opus 5.5 for coding agents and for any job that needs a score GPT-6 Sol does not reach at any effort setting, preferably at medium effort, where Cognition's FrontierCode has it ahead of every Sol setting for $0.80 a rollout; run GPT-6 Sol at xhigh for high-volume work that a mid-range score already clears, such as business automations, where Zapier has it level with Opus 5.5 at high for 38% of the cost per task.
claude-opus-5-5 / gpt-6-solRate card read Sep 24, 2026Updated Oct 1, 2026
OpenAIReleased Sep 22, 2026
GPT-6 Sol is a cheaper GPT-5.6 Sol: outside evaluators score it within a few points of its predecessor at roughly half the cost per task, and every leaderboard that lists both it and Claude Opus 5.5 ranks Opus 5.5 higher at Sol's best effort, so move to Sol to cut a bill, and run it at xhigh, which on Zapier's AutomationBench beats max for less.
gpt-6-solRate card read Sep 24, 2026Updated Oct 1, 2026
ComparisonAnthropicReleased Sep 22, 2026
Run coding, knowledge work and most agent jobs on Claude Opus 5.5, which Artificial Analysis scores above Claude Fable 5.1 at every effort setting but low for less per task, and move a job to Fable 5.1 only when Opus 5.5 at high or xhigh effort fails your own evaluations on long-running agent work, where one outside board, LMArena's Agent Arena, still ranks Fable first and has not yet listed Opus 5.5.
claude-opus-5-5 / claude-fable-5-1Rate card read Sep 23, 2026Updated Sep 24, 2026
ComparisonAnthropic and OpenAIReleased Sep 22, 2026
Send coding agents and knowledge work to Claude Opus 5.5 at medium or high effort, where Artificial Analysis scores it above GPT-6 Astra at the same setting for a similar or lower cost per task; keep GPT-6 Astra for jobs that need max effort at the lowest cost per task, where the outside numbers favour it.
claude-opus-5-5 / gpt-6-astraRate card read Sep 23, 2026Updated Sep 24, 2026
AnthropicReleased Sep 22, 2026
Claude Opus 5.5 is worth pricing at medium effort, where its savings are quoted and where Artificial Analysis scores it 51 and not the 58 in the headlines. Treat the launch post's safety lines as summaries of a system card that also reports a regression on text users paste into their own prompts, and read that card before routing third-party text through the model.
claude-opus-5-5Rate card read Sep 23, 2026Updated Sep 24, 2026
OpenAIReleased Sep 3, 2026
GPT-6 Astra ranks first on Epoch AI's capability index, first in LMArena's WebDev arena and second on Artificial Analysis v4.2 at max effort, at $10 and $50 per million tokens. GPT-5.6 Sol's $4 and $20 is promotional to 21 November 2026. Every headline number on this model moves with a setting, so price the workload you would actually move, at the tier you would actually run it.
gpt-6-astraRate card read Sep 5, 2026Updated Sep 6, 2026
