Frontier model prices in October 2026: what changed and what each one costs now

Between late September and mid-October 2026, Anthropic, OpenAI, Google and xAI all changed their API prices. The new catalogs settle into three bands. A $10 input and $50 output top tier for the hardest reasoning. A $2 input and $10 output middle tier that most engineering work now runs on. And a floor of $0.10 input and $0.50 output for classification and extraction. Opus 5.5, and Gemini 4 Argon from December, sit between the top and the middle at $4 and $20. All prices here are per million tokens.

What changed

• Anthropic cut Opus and Haiku. Claude Opus 5.5 costs $4 input and $20 output, against $5 and $25 for Opus 5. Claude Sonnet 5.5 arrived at $2 and $10. Claude Haiku 5.5, released on 7 October, costs $0.10 and $0.50 for prompts up to 100,000 tokens, a tenth of Haiku 4.5's $1 and $5. Claude Fable 5.1 keeps Fable 5's $10 and $50, but its cache reads drop from $1 to $0.25.
• OpenAI moved to the GPT-6 family. GPT-6.1 Sol costs $2 input and $10 output. Kapyn previously listed GPT-5.6 Sol at $4 and $20, so the middle of OpenAI's range is now half the price. GPT-6 Luna covers high volume at $0.10 and $0.50, and GPT-6 Astra holds the top at $10 and $50.
• Google launched Gemini 4 Argon with an end date on its price. Argon became available on 13 October. It costs $2 input and $10 output until 30 November, then $4 and $20 from 1 December.
• xAI priced Grok 4.7. Grok 4.7 costs $2 input and $6 output, with a 500,000 token context window. Its output rate is the lowest in the middle tier.

The October 2026 price table

Standard API rates for text, per million tokens, as listed on each vendor's pricing page on 21 October 2026:

Model                 Input   Output
Claude Fable 5.1     $10.00   $50.00
GPT-6 Astra          $10.00   $50.00
Claude Opus 5.5       $4.00   $20.00
Gemini 4 Argon *      $2.00   $10.00
Claude Sonnet 5.5     $2.00   $10.00
GPT-6.1 Sol           $2.00   $10.00
Grok 4.7              $2.00    $6.00
Claude Haiku 5.5 **   $0.10    $0.50
GPT-6 Luna            $0.10    $0.50

*  Until 30 Nov; $4 / $20 from 1 Dec.
** Prompts up to 100K tokens.

The vendors now match each other at almost every band. Fable 5.1 and Astra share the top price. Sonnet 5.5, Sol and Argon, for now, share the middle. Haiku 5.5 and Luna share the floor. On list price alone, there is little left to choose between the big labs within a tier, so quality on your own tasks and the details below decide it.

What a month of usage costs

Take a workload of 50 million input tokens and 10 million output tokens a month. That five-to-one ratio is typical of retrieval, review and summarisation, where the model reads far more than it writes. The monthly cost at each band:

• $10 / $50 (Fable 5.1, Astra): 50 × $10 + 10 × $50 = $1,000.
• $4 / $20 (Opus 5.5, Argon from December): 50 × $4 + 10 × $20 = $400. The same work on Opus 5 costs $500.
• $2 / $10 (Sonnet 5.5, Sol, Argon until 30 November): 50 × $2 + 10 × $10 = $200. At GPT-5.6 Sol's previous $4 and $20, it was $400.
• $2 / $6 (Grok 4.7): 50 × $2 + 10 × $6 = $160.
• $0.10 / $0.50 (Haiku 5.5, Luna): 50 × $0.10 + 10 × $0.50 = $10. On Haiku 4.5, it was $100.

The same month costs $1,000 at the top and $10 at the floor, a factor of 100. Moving routine steps, such as classifying, extracting and routing, off frontier models is still the largest saving available. Choosing between vendors inside a band moves the bill far less.

Three things the headline rates leave out

• Tokenizers differ. Simon Willison found that Haiku 5.5's tokenizer counts roughly 25 percent more tokens than Haiku 4.5 for the same text. In the example above, that turns $10 into about $12.50: still about 87 percent cheaper than Haiku 4.5, but short of the headline 90. Compare costs on your own prompts, not on token counts from another model.
• Long prompts cost more on Haiku 5.5. Above 100,000 tokens, Haiku 5.5 bills $0.50 input and $2.50 output, five times its standard rate. If your prompts carry long documents, check which side of that line they fall on.
• Caching changes the input bill. Cache reads cost $0.10 on Sonnet 5.5, $0.20 on Opus 5.5 and $0.25 on Fable 5.1. OpenAI charges $0.10 for cached input on Sol and $1.00 on Astra. Gemini 4 Argon's context caching costs $0.10 until 30 November and $0.20 after. For agents and chat products that resend the same long system prompt, cached input can matter more than the base rate.

Three takeaways

1. $2 and $10 is the new default middle tier. Sonnet 5.5 and Sol both sit there, and Argon does until December. If you last priced a workload on GPT-5.6 Sol, the same work now costs about half.
2. Argon doubles on 1 December. At $2 and $10 it matches Sonnet 5.5 and Sol. From 1 December it costs $4 and $20, the same as Opus 5.5. If you pick Argon for price, budget for the December rate.
3. The floor is low enough to stop rationing. At $0.10 and $0.50, Haiku 5.5 and Luna make high-volume classification and extraction a rounding error next to the rest of a stack. Check the tokenizer and the 100,000 token line before you rely on the exact figure.

Sources

• Anthropic: Claude API pricing
• OpenAI: API pricing
• Google: Gemini API pricing
• Google: Introducing Gemini 4 Argon
• xAI: Models and pricing
• Simon Willison: notes on Claude Haiku 5.5

Kapyn's model comparison tracks the current version of each model family. For choosing between the two biggest labs, see Claude vs GPT.