Anthropic's Claude pricing docs now give Claude Haiku 5.5, dated October 7, 2026, a $0.01 cache-read rate for prompts up to 100K tokens. Per million tokens, Opus 5.5 costs $4 in and $20 out with $0.20 cache reads, GPT-6 Sol and 6.1 Sol cost $2 and $10, and Luna costs $0.10 and $0.50 with $0.01 cache reads.
What changed: Haiku 5.5 and Sonnet 5.5 are priced, and GPT-6.1 Sol halved Sol's cache rate. Who's affected: Agent builders who pick a model per workload. What to do: Re-price cache-heavy loops and check prompt lengths against each long-prompt rule.
What changed since launch?
Anthropic's Opus 5.5 page still says Sonnet 5.5 and Haiku 5.5 "will follow in the coming weeks," but Anthropic dates Sonnet 5.5 to September 28, 2026 and Haiku 5.5 to October 7. The GPT-6.1 Sol model docs keep Sol's $2 input and $10 output, and OpenAI says cached input is "50% less than GPT-6 Sol's cached input pricing." So Opus 5.5 and Sol share a $0.20 cache rate only on the older GPT-6 Sol.
What does each model cost now?
Six tiers compete, and two pairs tie exactly. Standard rates per million tokens, checked October 21, 2026:
| Model | Input | Output | Cache read | Cache write | Long prompts |
|---|---|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | $0.20 | $5 (5-min) | Standard price across the 1M window |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 | $2.50 (5-min) | Standard price across the 1M window |
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.01 | $0.125 (5-min) | Over 100K: $0.50 in, $2.50 out, $0.05 cache |
| GPT-6.1 Sol | $2 | $10 | $0.10 | $2.50 | Over 272K: 2x input and cache, 1.5x output |
| GPT-6 Sol | $2 | $10 | $0.20 | $2.50 | Same 272K rule |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 | $0.125 | Same 272K rule |
Anthropic counts all input tokens toward Haiku's 100K line, "including cache reads and cache writes." OpenAI's multipliers apply "for the full request" past 272K tokens. Batch is half price at both vendors.
Is 40% cheaper the same as 50% cheaper?
No. Anthropic says Opus 5.5 "will cost 40% less than Opus 5 on typical workloads," while OpenAI says it cut Sol and Luna prices "by 50% compared with their GPT-5.6 promotional pricing." Opus 5.5's input and output prices fell 20%, from $5 / $25 to $4 / $20, and cache reads fell 60%, from $0.50 to $0.20. The 40% is Anthropic's blend of these on its own test workloads.
OpenAI's 50% is a list-price ratio: GPT-6 Sol's $2 / $10 is half of GPT-5.6 Sol's promotional $4 / $20. Don't set the two side by side.
Whose benchmarks can you trust?
Neither vendor's page on its own. Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 at xhigh effort, against 57.9% for GPT-6 Astra as reported by OpenAI. On AutomationBench, Zapier ran Opus 5.5 to 40.0% with no fallback models, so safeguard interventions counted as failures. OpenAI shows GPT-6 Sol (xhigh) at 33.2% for $0.27 per task and Opus 5 (max) at 26.9%. Anthropic's own table puts Astra at 41.4%, just above Opus 5.5.
The rows both pages share agree. Each vendor picks which rows and effort levels to show, and neither page has GPT-6.1 Sol, Sonnet 5.5 or Haiku 5.5. Test on your own traces.
Are you running the model you picked?
Not always. Anthropic says most cybersecurity tasks on Opus 5.5 "will be re-routed to Opus 4.8," with safeguards that "fall back to another model transparently." In its benchmark runs, biology and frontier LLM development tasks were completed by Opus 5.
Preserved thinking matters if you switch models. For API accounts created on or after August 31, 2026, the API checks requests to Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. Edit what comes before a thinking block and it returns a 400 error or drops the invalid blocks, whichever you choose. Opus 5.5 also can't run with thinking switched off.
Which should you switch to?
In an agent loop, the cache-read rate is the line that repeats, and every tier now has a same-priced rival. Our agent cost guide shows how to break a bill down that way.
- Cheap tier: Haiku 5.5 and GPT-6 Luna both charge $0.01 for cache reads, 20x below Opus 5.5. From 100K to 272K tokens, Haiku's rises to $0.05 while Luna stays at $0.01.
- Mid tier: Sonnet 5.5 and GPT-6.1 Sol match on every rate. Past 272K tokens, Sol rises to $4 in, $0.20 cache and $15 out; Sonnet stays at $2, $0.10 and $10.
- Opus 5.5: its $0.20 cache read is 2x GPT-6.1 Sol's, so pay it only when your evals show it earns it. Our LLMOps explainer covers tracking that, and the agent frameworks comparison covers switching costs.
- Replaying or editing thinking blocks: migrate before moving to any Claude 5.5 model.
- On GPT-5.6 Sol's promotional rate: OpenAI says it runs "at least through November 21, 2026," so plan your move before then.
What is still unconfirmed?
- GPT-6.1 Sol's release date: the OpenAI pages we read give none.
- GPT-5.6 Sol's price after November 21, 2026: check the vendor page.
- Quality: all benchmark figures are vendor-reported, and Toolbit did not test these models.
