Opus 5.5 vs GPT-6 Astra vs Gemini 4 Argon: price, specs

Quick answer: per token, Gemini 4 Argon is the cheapest of the three until 30 November 2026, at $2 per million input tokens and $10 per million output tokens. From 1 December it costs $4 / $20, the same as Claude Opus 5.5. GPT-6 Astra costs $10 / $50, two and a half times Opus 5.5, and OpenAI charges more again once a prompt passes 272K input tokens. All three have a context window of about a million tokens. Opus 5.5 and Astra return up to 128K output tokens per request; Argon returns 65,536 by default and up to 1,048,576 with long decode. The benchmark claims below are each vendor's own, so test on your own work before you switch.

Key facts

• Claude Opus 5.5 (claude-opus-5-5, released 22 September 2026): $4 input / $20 output per million tokens, $0.20 cache hits, 1M-token context, 128K max output, medium effort by default.
• GPT-6 Astra (gpt-6-astra, in the API since 3 September 2026): $10 / $50, $1 cached input, 1,050,000-token context with up to 922,000 input tokens, 128K max output.
• Gemini 4 Argon (gemini-4-argon, broadly available since 13 October 2026): $2 / $10 through 30 November, then $4 / $20. 1,048,576 tokens shared by input and output.
• Long prompts: Anthropic bills the full 1M context at standard rates, OpenAI adds a surcharge above 272K input tokens, and Google lists no prompt-length tiers for Argon.
• Batch requests cost half price on all three.
• Only Argon accepts video and audio input. Opus 5.5 and Astra take text and images.

Price per million tokens

• Claude Opus 5.5: $4 input, $20 output, $0.20 cache hits, $5 for 5-minute cache writes. Batch: $2 / $10. Fast mode: $8 / $40.
• GPT-6 Astra: $10 input, $50 output, $1 cached input, $12.50 cache writes. Batch and Flex: half price. Fast mode: double.
• Gemini 4 Argon until 30 November: $2 input, $10 output (thinking tokens count as output), $0.10 implicit caching. From 1 December: $4, $20 and $0.20. Batch: $1 / $5, then $2 / $10. There's no free tier.

A worked example without caching: a request with 100,000 input tokens and 10,000 output tokens costs $0.60 on Opus 5.5, $1.50 on Astra and $0.30 on Argon until 30 November ($0.60 after). With 400,000 input tokens, Astra's long-prompt rates kick in, and the same request costs $1.80 on Opus 5.5, $8.75 on Astra and $0.90 on Argon ($1.80 after November). Real bills also depend on how many tokens each model spends thinking.

Context window and output

• Opus 5.5: 1M-token context and 128K max output on the Messages API. On the Batch API it can return up to 300K output tokens with the output-300k-2026-03-24 beta header.
• GPT-6 Astra: 1,050,000-token context, of which up to 922,000 can be input, and 128,000 max output.
• Gemini 4 Argon: 1,048,576 tokens shared between input and output. Synchronous requests return up to 65,536 output tokens by default, and long decode raises that to 1,048,576. It takes text, images, video, audio and PDFs.

Long prompts: only Astra charges extra

Anthropic's pricing page says Opus 5.5 includes the full 1M-token window at standard prices, so a 900K-token request pays the same per-token rate as a 9K one. OpenAI's Astra page bills the whole request at 2x the input and cache rates and 1.5x the output rate once input passes 272K tokens. Google's Gemini API pricing lists one rate for Argon at any prompt length, unlike Gemini 3.1 Pro Preview, which costs more above 200K tokens.

Benchmarks: vendor-reported only

There's no head-to-head test from a neutral source here. Each lab reports its own numbers with its own settings, so read these as claims, not a ranking:

• Anthropic: Opus 5.5 scores 66.4% on Terminal-Bench 4.0 at xhigh effort. Its table lists GPT-6 Astra at 57.9%, OpenAI's own figure at high effort, and shows Astra ahead on AutomationBench (41.4% against 40.0%).
• Google: Argon scores 77.9% on DeepSWE v1.1 and 51.3% on Zapier's AutomationBench, per its 30 September launch post, which doesn't compare it with Opus 5.5 or Astra.
• Independent results from open datasets (Arena, Epoch AI) are on our benchmarks page.

Opus 5.5 vs Fable 5.1

Claude Fable 5.1 is Anthropic's model for demanding reasoning and long-horizon agentic work. It costs $10 / $50 per million tokens with $0.25 cache hits, against $4 / $20 and $0.20 for Opus 5.5. Both have a 1M-token context and 128K output, but Fable defaults to high effort, Opus 5.5 to medium, and Anthropic rates Fable as slower. Anthropic's advice, from its models overview:

Anthropic suggests Fable 5.1 when your evals on Opus 5.5 at higher effort still fall short, and it claims Opus 5.5 performs at Fable 5.1's level on most work. In the Claude apps, Opus is on every paid plan, while Fable needs usage credits on Pro and is limited to 50% of weekly limits on Max and on Team Premium seats.

Opus 5 vs Opus 5.5

• Price: $5 / $25 for Opus 5, $4 / $20 for Opus 5.5, 20% less. Cache hits fell from $0.50 to $0.20.
• Fast mode: $10 / $50 on Opus 5 (research preview), $8 / $40 on Opus 5.5.
• Default effort: high on Opus 5, medium on Opus 5.5. Set effort explicitly when you migrate.
• Thinking: Opus 5 still lets you disable it at high effort or below. On Opus 5.5 it's always on, and forced tool use returns an error.
• Knowledge cutoff: May 2026 for Opus 5, June 2026 for Opus 5.5.
• Anthropic's claims: Opus 5.5 generates output more than 30% faster than Opus 5 and costs 40% less to run.
• Opus 5 (released 24 July 2026) is now a legacy model. It's still available, with retirement not sooner than 24 July 2027.

Where you can use them

• Opus 5.5: the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. In the Claude apps it's on Pro and every higher plan, not Free.
• GPT-6 Astra: the OpenAI API and Codex. OpenAI estimates 5 to 45 local Codex messages per five hours on Plus and Business, and Pro currently has no five-hour limit.
• Gemini 4 Argon: the paid tier of the Gemini API, in Google AI Studio and Android Studio, and the Gemini app for Google AI Pro and Ultra subscribers.