Opus 5.5 and GPT-6 Sol cut the price per token but not the bill

On Tuesday, September 22, mid-morning in California, Anthropic released Claude Opus 5.5. OpenAI answered the same morning with two models of its own, GPT-6 Sol and GPT-6 Luna¹. Both announcements promised more capable models, as announcements do. What both companies led with, though, was the price.

Opus 5.5 costs $4 per million input tokens² and $20 per million output tokens, 20% less than Opus 5³. GPT-6 Sol drops to $2 and $10, half the price of GPT-5.6 Sol, and Luna goes for 10 and 50 cents⁴. The race for performance now comes with a price war on the side. That war is being fought one floor below the top, though, and the price of a token says surprisingly little about what the customer ends up paying.

Big-brother performance at a little-brother price

The two announcements make the same pitch, a flagship's performance at a mid-range price. Anthropic says Opus 5.5 "performs at the level of Claude Fable 5.1 on most work" and costs 40% less than Opus 5 to run on typical workloads³. Its own numbers even put it ahead of Fable on agentic coding, 66.4% on Terminal-Bench 4.0 against 55.8%⁵. OpenAI, for its part, says Sol makes about half as many mistakes as its predecessor, "reaching Astra-level reliability at much lower cost"¹.

OpenAI's favorite comparison goes straight at the competition. On AutomationBench, Sol reportedly beats Claude Opus 5 at 11.1 times lower cost, and on OSWorld 2.0, which measures how well a model operates a computer, it matches Opus 5 for 80% less⁶. The rival in those charts is Opus 5, which Anthropic had replaced that very morning. Apparently a pricing deck takes longer to build than a model.

OpenAI has been cutting for a while. In July, GPT-5.6 Sol launched at $5 input and $30 output, and I wrote back then that OpenAI was going after the king by way of the wallet. Since then, the same model has dropped to $4 and $20, a promotional rate guaranteed through at least November 21⁷. The new Sol lists at $2 and $10, and OpenAI says that price is permanent⁸. The output price of OpenAI's headline model has been cut by two-thirds in about ten weeks.

The top floor didn't move a cent

Look one floor up and the war runs out of ammunition. Claude Fable 5.1 and GPT-6 Astra, the two flagships and the most expensive models you can buy, still cost $10 per million input tokens and $50 output, and they're the only ones the September 22 fight left untouched⁹. Above them, Mythos and the full version of Astra remain locked away, as I wrote in early September. Even Opus 5.5 comes on a leash. When its safeguards trip on a cybersecurity or biology task, the request gets rerouted to Opus 4.8, an older model³.

So the performance race hasn't stopped. It carries on at the top, at full price and behind a door, while the floor below sells off the level the top reached a few weeks ago. This is also Anthropic's first release since Dario Amodei called for pacing the frontier, and it fits his essay neatly, since it pushes no frontier and only makes what already exists cheaper. Slowing down the top and running a clearance sale underneath go together just fine, especially for a company getting ready to go public¹⁰.

The token got cheaper, and the bill depends on you

If you pay by usage, the sticker price matters less than how many tokens a model burns to finish a job. That number depends on a setting you choose, the reasoning effort¹¹. Artificial Analysis, which runs every model through the same battery of tests, puts the same GPT-6 Sol at $0.13 per task on low effort and $1.06 on max effort, eight times as much¹². Opus 5.5 produced 260 million output tokens to get through those tests, against a median of 88 million for the models tested¹³. A cheaper token from a chatty model doesn't necessarily make a smaller bill.

Anthropic's 40% figure deserves the same scrutiny. It assumes moving from Opus 5's default setting, high effort, to Opus 5.5's, which is medium¹⁴. Part of the savings comes from a model that thinks for less time unless you ask it to, which may or may not be enough for the job at hand. Simon Willison, a developer who puts every new model through its paces, watched his max-effort runs of Opus 5.5 cost $2.56 each, take 20 minutes, and hit the output ceiling before finishing⁹.

The most useful cut hides on a quieter line. Rereading context that's already cached¹⁵ now costs 20 cents per million tokens at Anthropic, down from 50 with Opus 5³. An agent that rereads the same codebase a hundred times in a session pays mostly for that line, far more than for the price in the headline.

Subscribers get more, not cheaper

Neither announcement touches subscription prices. Anthropic is raising the five-hour usage limits on its Pro, Max, and Team plans and giving every subscriber one limit reset to spend whenever they like³. OpenAI is opening Sol to Plus subscribers and up, and Luna to free accounts in its desktop app⁴. Subscribers get more for the same money, while the actual price cuts go to the companies and developers who pay by the token.

I'm one of those subscribers. My Anthropic plan costs exactly what it cost on Monday, and the price war buys me a little more room before I hit the ceiling. I'm not complaining.

Chinese labs were already slashing prices

The price war didn't start in California. DeepSeek charges $1.32 per million input tokens and $3.96 output for its V4 Pro at peak hours, and half that the rest of the time, which still undercuts the new Sol¹⁶. Chinese open models even fit on a gaming graphics card, with Qwen leading the pack. With Luna, OpenAI slips under some of them, 29% below Xiaomi's MiMo on input price⁸. There's some irony here, since the Chinese labs learned partly by copying Claude. The American labs are now competing on price with rivals they trained without meaning to.

That leaves a paradox neither OpenAI nor Anthropic really explains. Memory has never been this expensive, the chipmakers have already sold their 2027 output, and the price per token keeps falling anyway. OpenAI credits improvements in caching and inference¹. Part of the answer is probably the competition itself. When two companies launch on the same morning, neither can afford to keep last month's prices.

Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks³, so the next round will be fought even further down the lineup. The day I'm waiting for is the one when a Fable or an Astra drops below $10. That's when the performance race will really have slowed down. Until then, both rivals are selling last month's best at a discount and keeping the rest under lock and key.

Notes

1. TechCrunch, "OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes", September 22, 2026. ↩ ↩² ↩³

2. A token is the unit of text a model reads and writes, a word or a piece of one. Labs bill by the million tokens, and output costs more than input, since writing text takes more computation than reading it. ↩

3. Anthropic, "Introducing Claude Opus 5.5", September 22, 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

4. OpenAI, "Introducing GPT-6 Sol and Luna", September 22, 2026. ↩ ↩²

5. VentureBeat, "Anthropic releases Claude Opus 5.5, beating Fable 5.1 on key agentic benchmarks at 60% cheaper API price", September 22, 2026. ↩

6. MarkTechPost, "OpenAI Releases GPT-6 Sol and Luna", September 22, 2026. The measurements are OpenAI's own. ↩

7. OpenAI, GPT-5.6 Sol model page and GPT-6 Sol model page, accessed September 23, 2026. ↩

8. VentureBeat, "OpenAI releases GPT-6 Sol and Luna models, slashing API costs 50% or more", September 22, 2026. ↩ ↩²

9. Simon Willison, "Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war", September 22, 2026. ↩ ↩²

10. Bloomberg, "Anthropic Unveils More Cost-Efficient Opus 5.5 Model Before IPO", September 22, 2026. ↩

11. Reasoning effort sets how long the model thinks before it answers. The longer it thinks, the more tokens it writes that nobody ever reads and everybody pays for. ↩

12. Artificial Analysis, GPT-6 Sol page, accessed September 23, 2026. ↩

13. Artificial Analysis, Claude Opus 5.5 page, accessed September 23, 2026. ↩

14. Digital Applied, "Claude Opus 5.5: Pricing, Benchmarks and Breaking Changes", September 2026. ↩

15. The cache keeps text the model has already read, a long document or a whole codebase, so it can reread it without processing everything again. Labs bill that reread at a fraction of the price of a fresh read. ↩

16. DeepSeek, API pricing, accessed September 23, 2026. Peak hours fall on weekdays, 1 to 4 a.m. and 6 to 10 a.m. UTC, which is the middle of the night in the United States. ↩