Pricing snapshot · September 2026
Claude vs. GPT-6 API pricing, September 2026
Two pairs of models carry identical list prices. Only their cache-read prices differ, and the cheaper provider changes by tier.
01 / Fair pairs
Compare models at the same list price.
Claude Fable 5.1 and GPT-6 Astra both list at $10 per million input tokens and $50 per million output tokens. Claude Sonnet 5.5 and GPT-6.1 Sol both list at $2 and $10. Those are the fair head-to-head pairs.
Claude Opus 5.5, at $4 and $20, has no GPT-6 model at its price. Compare it with Astra only on a task where both meet your quality floor, and then compare cost per successful outcome. Haiku 4.5 and Luna sit an order of magnitude apart and serve different jobs.
02 / Rate card
Standard API rates per million tokens.
| Model | Input | Cache read | Cache write | Output |
|---|---|---|---|---|
| Claude Fable 5.1 | $10.00 | $0.25 | $12.50 | $50.00 |
| GPT-6 Astra | $10.00 | $1.00 | $12.50 | $50.00 |
| Claude Opus 5.5 | $4.00 | $0.20 | $5.00 | $20.00 |
| Claude Sonnet 5.5 | $2.00 | $0.20 | $2.50 | $10.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $2.50 | $10.00 |
| Claude Haiku 4.5 | $1.00 | $0.10 | $1.25 | $5.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 | $0.50 |
Anthropic's cache prices reflect its 5-minute cache lifetime. OpenAI lists “cached input” for the read rate. Above 272K input tokens, OpenAI doubles the input and cache rates for all three GPT-6 models and raises output by half, so Astra's output goes from $50 to $75. Both providers list a 50% batch discount.
03 / The difference
Only the cache-read price differs, in opposite directions by tier.
Within each matched pair, input, output, and cache-write prices are identical. Only the cache-read price differs:
- Flagship tier: Fable 5.1 reads cache at $0.25 per million tokens; Astra at $1.00. Astra costs 4 times as much per cached token.
- Mid tier: Sonnet 5.5 reads cache at $0.20; Sol at $0.10. Sonnet costs twice as much per cached token.
Neither provider is cheaper across the board. The gap grows with the cached share of each request, which runs high in agent loops, long system prompts, and tool-heavy workflows that resend the same context on each call. With little caching, the matched pairs cost the same per token.
All four matched models break even on a cache write after 2 reuses by the formula in the prompt-caching break-even guide, so the write price does not separate them either.
04 / Illustrative workload
100,000 agent calls with 90% of input cached.
Take 100,000 requests, each with 50,000 input tokens (45,000 read from an already-written cache, 5,000 new) and 1,000 output tokens. The cached column assumes the prefix stays warm; the uncached column bills all 50,000 input tokens at the input rate.
| Model | 90% cached | No caching |
|---|---|---|
| Claude Fable 5.1 | $11,125 | $55,000 |
| GPT-6 Astra | $14,500 | $55,000 |
| Claude Opus 5.5 | $4,900 | $22,000 |
| Claude Sonnet 5.5 | $2,900 | $11,000 |
| GPT-6.1 Sol | $2,450 | $11,000 |
| Claude Haiku 4.5 | $1,450 | $5,500 |
| GPT-6 Luna | $145 | $550 |
Without caching, Fable 5.1 and Astra cost the same $55,000. With 90% of input cached, Astra costs $14,500 against $11,125, about 30% more. In the mid tier the direction reverses: Sonnet 5.5 costs $2,900 against Sol's $2,450, about 18% more.
Fable 5.1 per request: 5,000 × $10 + 45,000 × $0.25 + 1,000 × $50, per million tokens, is $0.11125. Astra: 5,000 × $10 + 45,000 × $1 + 1,000 × $50 is $0.145. The workload is hypothetical and shows how the rates interact. Your own token mix sets your bill.
05 / Beyond the rate card
Identical per-token prices can still produce different costs per task.
- Tokenizers differ. The same prompt produces a different token count on each provider, so identical rates can still produce different charges for the same text. The gap can be large: Anthropic's pricing documentation states that Claude 4.7 and later models produce about 30% more tokens for the same text than earlier Claude models. Measure token counts on your own prompts.
- Tokens per task differ. A model that finishes in fewer turns or shorter outputs costs less per task at the same rate. Vendor claims about this need your own measurement.
- Success rates differ. A cheaper call that fails more often can cost more per successful outcome. See reducing costs without losing quality.
- Retries multiply everything. Every rate above applies per attempt. See LLM retry costs.
Put two models into the same cost-per-outcome comparison.
The free calculator adds retries, success rate, and review labor to the token rates.
06 / Common questions
Quick answers.
Is Claude cheaper than GPT-6?
At the flagship tier with heavy caching, Fable 5.1 costs less than Astra. At the mid tier with heavy caching, Sol costs less than Sonnet 5.5. With little caching, the matched pairs cost the same per token.
Is Opus 5.5 vs. GPT-6 a fair comparison?
Not on price alone. No GPT-6 model lists at Opus 5.5's $4 and $20. Compare it against Astra or Sol only on a task where each meets your quality floor, using cost per successful outcome.
Does batch pricing change the comparison?
Both providers list a 50% batch discount, so the ratios between matched models stay the same for batch-eligible work.
Will this page be updated when prices change?
No. It records the published rates for these model versions on September 30, 2026. Check the provider pages below for current rates.
07 / Sources
Checked against the providers' own pricing pages.
Rates were read from Anthropic's pricing page and OpenAI's API pricing page on September 30, 2026. decost.ai receives no compensation from either provider; see the method and independence disclosure.
Corrected September 30, 2026: an earlier version called OpenAI's mid-tier model GPT-6 Sol; its pricing page lists it as gpt-6.1-sol. It also said Astra's long-context output rate doubles; it rises from $50 to $75. No other figure changed.