Price cuts minutes apart send agent routing back to the spreadsheet
Anthropic took 20% off Opus 5.5 and OpenAI halved its two new GPT-6 tiers the same day. The deepest cuts landed on cached input reads, so what any pipeline actually saves depends on its cache hit rate.
The Product Desk
What happened
- Anthropic released Claude Opus 5.5 with a 20% price cut, at $4 per million input tokens and $20 per million output, and a faster serving mode at double both rates.
- Minutes later OpenAI put out two GPT-6 models, Sol at $2 and $10 per million and Luna at 10 and 50 cents, half the price of the GPT-5.6 versions that carried the same names.
- Anthropic put Opus 5.5 at 66.4% on the agentic coding test Terminal-Bench 4.0 against 55.8% for Fable 5.1, and said output generation is more than 30% faster than Opus 5.
- Anthropic said attempts to circumvent containment boundaries fell 85% against Opus 5, and that Frontier Design and METR both tested the model before release.
- Opus 5.5 is live on the Claude Developer Platform and through AWS, Google Cloud and Azure, with Claude Sonnet 5.5 and Haiku 5.5 due in the coming weeks.
Why it matters
- Who has to decide decision Any router configured on last quarter's rates now has the wrong tier winning: the gap between the two vendors' output prices is a clean 2x, and it was set by two announcements on the same afternoon.
- What it limits constraint OpenAI published a cost per task and Anthropic published completion rates, so a buyer cannot build cost per completed task for both sides from the launch figures without measuring tokens per attempt itself.
- What it costs cost The saving is not evenly distributed: teams whose agents replay long prompts collect most of it through cached reads, and teams generating long fresh answers collect the posted cut and little else.
- Who's exposed exposure Anyone who priced a security-facing agent on Opus 5.5 rates is exposed, because the tasks that matter get served by Opus 4.8.
Next story · Build · Oct 7
Prompt length sets what Anthropic's Haiku 5.5 price cut is worth to each workloadConfirmed22 publishers
Cache reads on Opus were 50 cents per million before this week. They are 20 cents now. OpenAI's discount on reads of cached input tokens is 90%, and against Sol's $2 input list that also works out to 20 cents per million; on Luna it is a penny. OpenAI said coding agents can now change reasoning effort and the set of tools available to them without breaking the cache. A broken cache turns a 20-cent read back into a $2 one.
Anthropic gave the saving in two sizes. Posted prices came down 20%, and the company said a typical workload lands 40% below Opus 5. The extra 20 points come from the shape of the job, so a pipeline with little reuse and long outputs gets the 20% and stops there. Cache reads fell furthest, and the answers are shorter: Yashodha Bhavnani, vice president of AI products at Box, reported answers 40% less verbose without losing accuracy, in testing notes Anthropic published with the release.
OpenAI put Sol at 33.2% of AutomationBench tasks at 27 cents apiece, which is about 81 cents per completed task. Anthropic reported 40% completion on the same test for Opus 5.5 against 26.9% for Opus 5, and did not publish a cost per task. Opus 5.5 output is $20 per million against Sol's $10. Run both at the same token count per attempt and an Opus attempt costs 54 cents, or $1.35 per completed task. To match Sol's 81 cents, Opus 5.5 has to finish an attempt on 60% of the tokens Sol uses. Mario Rodriguez, GitHub's chief product officer, said that in VS Code the model "solved more terminal tasks than Opus 5 in less than half the steps." Steps are not tokens, and that comparison is against Opus 5.
What you pay also depends on the task. Restrictions on cybersecurity work carry over from Fable 5.1, and most such tasks are routed to the older Opus 4.8 model. High-risk biology work is fenced off behind verification programs that now include a life sciences track.
The routing comes down to what share of your input tokens are cache reads and what a failed attempt costs you. Low reuse with an expensive failure, such as a code migration a person has to sign off, is where paying twice Sol's output rate can work out, if the step count holds outside a VS Code demo. High reuse with a cheap retry, such as the summarization and extraction jobs you can run twice without a human noticing, points the other way, to Luna, where an output token costs 40 times less than one from Opus 5.5.
What to watch
- Third-party runs of Opus 5.5 against Sol on one harness; both labs benchmarked against Opus 5 and Fable 5.1 instead.
- Whether Anthropic's cybersecurity routing to Opus 4.8 moves once the cheaper Claude tiers ship.
- OpenAI's claim that Luna at higher effort settings matches GPT-5.6 Sol at roughly a hundredth of the cost, tested on real extraction work.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption20
- Hype gap+25
- Incentives70
- Confidence60
Perspective Coverage
5 publishers- Builder
- Builder 38%
- Operator
- Operator 37%
- Investor
- Investor 25%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Anthropic released Claude Opus 5.5 and cut its price 20%.
- [2]
Input on Opus 5.5 costs $4 per million tokens and output $20 per million.
- [3]
Anthropic said the new Opus 5.5 pricing works out to 40% less on a typical workload than Opus 5.
- [4]
Cache reads on Opus 5.5 took the sharpest cut, dropping 60% to 20 cents per million tokens.
- [5]
Minutes after the Opus 5.5 release, OpenAI put out two new GPT-6 models, Sol and Luna, at half what their predecessors cost.
- [6]
OpenAI halved the price of both new models against the GPT-5.6 versions that carried the same names, bringing Sol to $2 per million input tokens and $10 per million output.
- [7]
Luna is priced at 10 cents per million input tokens and 50 cents per million output.
- [8]
Terminal-Bench 4.0, an agentic coding test, came back at 66.4% for Opus 5.5 against 55.8% for Fable 5.1.
- [9]
Output generation on Opus 5.5 is more than 30% faster than Opus 5, which shipped in late July.
- [10]
On AutomationBench, Opus 5.5 reported 40% task completion where Opus 5 managed 26.9%.
- [11]
Opus 5.5 is available on the Claude Developer Platform as claude-opus-5-5 and through Amazon Web Services, Google Cloud and Microsoft Azure, with Claude Sonnet 5.5 and Claude Haiku 5.5 due in the coming weeks.
- [12]
OpenAI is discounting reads of cached input tokens by 90%.
- [13]
OpenAI said cache hit rates have improved, and that coding agents can now change reasoning effort and the set of tools available to them without breaking the cache.
- [14]
Anthropic said attempts to circumvent containment boundaries dropped 85% against Opus 5, and that Frontier Design and METR both tested the model before release.
- [15]
Restrictions on cybersecurity work carry over from Fable 5.1, and most such tasks are routed to the older Opus 4.8 model.
- [16]
Biology work judged high risk is fenced off, with broader access running through verification programs that now include a life sciences track.
- [17]
OpenAI's comparisons run against Fable 5.1 and Opus 5, so neither launch offers a direct head-to-head against the other.
- [18]
A faster Opus 5.5 serving mode carries a premium, at $8 per million input tokens and $40 per million output.
- [19]
Before the cut, Opus cache reads cost 50 cents per million tokens.
- [20]
OpenAI reported Sol completing 33.2% of tasks on AutomationBench at 27 cents apiece.
- [21]
Yashodha Bhavnani, vice president of AI products at Box Inc., said answers from the new model were 40% less verbose without losing accuracy, in testing notes Anthropic published with the release.
- [22]
GitHub Inc. Chief Product Officer Mario Rodriguez said that in VS Code the model "solved more terminal tasks than Opus 5 in less than half the steps."
- [23]
OpenAI said Luna at higher effort settings matches GPT-5.6 Sol at roughly a hundredth of the cost.
- [24]
Opus 5.5 output tokens cost twice Sol's rate, $20 per million against $10.
- [25]
If Opus 5.5 spent the same tokens per AutomationBench attempt as Sol, an attempt would cost 54 cents and a completed task about $1.35 at 40% completion.
- [26]
To match Sol's 81 cents per completed task, Opus 5.5 must complete an attempt on 60% of the tokens Sol uses.
- [27]
An Opus 5.5 output token costs 40 times a Luna output token, $20 per million against 50 cents.
- [28]
With a 90% discount on Sol's $2 per million input price, a cached read on Sol costs 20 cents per million.
- [29]
With the same 90% discount, a cached read on Luna costs 1 cent per million input tokens.
- [30]
At 27 cents per attempt and 33.2% completion, Sol costs about 81 cents per completed AutomationBench task.
Sources
5 independent publishers whose own reporting we read for this story.
- 9to5google.comClaude Opus 5.5 and OpenAI GPT-6 Sol & Luna both launch today with lower costs
2 articles · September 22, 2026
- arstechnica.comNew Anthropic, OpenAI models make same promise: A little more for a lot less money
2 articles · September 22, 2026
- cnet.comAnthropic and OpenAI Drop New High-Efficiency Models - CNET
1 article · September 22, 2026
- engadget.comAnthropic and OpenAI announce more powerful (and cheaper) AI models
2 articles · September 22, 2026
- siliconangle.comAnthropic releases Claude Opus 5.5 and OpenAI counters with two cheaper GPT-6 models
2 articles · September 22, 2026