TL;DR
- Anthropic shipped Claude Opus 5.5 on September 22 at $4 input / $20 output per million tokens ā 20% below Opus 5 and about 40% cheaper per typical workload, with cache reads at $0.20/M.
- Ninety minutes later, OpenAI launched GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50, a flat 50% cut against GPT-5.6 promotional pricing.
- Output pricing now spans roughly 100x inside one model family: GPT-6 Astra at $50/M against Luna at $0.50/M.
- The benchmarks disagree (AutomationBench: Opus 5.5 at 40.0%, Sol at 33.2%), but the cost-per-task gap is 11x. Per-task cost, not per-token price, is what decides which model an agent runs.
Two launches, 90 minutes apart
Frontier launches are usually capability stories. Tuesdayās were pricing stories. Anthropic opened with Claude Opus 5.5, pitched as Fable 5.1-class performance at 40% less cost than Opus 5 on typical workloads: $4/$20 per million input and output tokens, cache reads at $0.20 per million (60% cheaper), output more than 30% faster, and higher five-hour usage caps. (Source : Anthropic ā Introducing Claude Opus 5.5)
Ninety minutes later, OpenAI answered with GPT-6 Sol and Luna and a flat 50% cut versus GPT-5.6 promotional rates. Sol lands at $2/$10 ā half of Opus 5.5 ā while Luna sits at $0.10/$0.50, or 2.5% of Opus 5.5 and 1% of Astraās $10/$50. OpenAI says caching now delivers 90% discounts on cached input reads. (Source : OpenAI ā Introducing GPT-6 Sol and Luna)
Both labs had signalled the shift: Anthropic published a āpacing the frontierā argument and then shipped a model that is cheaper and stronger, while OpenAI had already positioned Astra as its expensive tier. We covered that framing in the Astra launch analysis and in the earlier GPT-5.6 Sol/Terra/Luna pricing breakdown.
What the benchmarks actually say
Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0, 57.8% on CursorBench 4.0, 1846 Elo on GDPval-AA v2.1, 40.0% on Zapierās AutomationBench and 81.8% partial credit on OSWorld 2.0, plus a 680,000-line migration completed in under a day by one tester. (Source : Anthropic ā Opus 5.5 System Card)
OpenAIās numbers are more about cost: Sol at xhigh effort scores 33.2% on AutomationBench at $0.27 per task, against Claude Opus 5 at max effort on 26.9% and 11.1x that cost, plus 56.4% on Agentsā Last Exam at 60% lower cost per task. Open weights make the same argument: DeepSeekās V4.1-Flash is a 552B-parameter mixture-of-experts with 8B active parameters on input and 16B on output, and a KV cache needing a quarter of the HBM and an eighth of the SSD footprint of the previous generation. (Source : DeepSeek ā Introducing DeepSeek-V4.1-Flash)
No clean same-harness comparison exists. Anthropic benchmarks against Astra and GPT-5.6 Sol; OpenAI mostly compares to older Claude models. Third-party probes split too: Browser Use measured Sol at 66.9 versus Opus 5.5 at 59.4 on its browser benchmark with Sol about 3.5x cheaper, Artificial Analysis ranked Opus 5.5 Max first on its Intelligence Index at 58, and a 10-task practitioner run preferred Opus on seven tasks at $213 and 8h40 versus Solās $74 and 5h51. (Source : The Neuron ā Everything That Happened in AI Today, September 22, 2026)
Why it got cheap
Three levers, only one of them a price cut. Caching first: agentic work is dominated by re-sent context, so cache read pricing decides the bill. Architecture second: DeepSeekās asymmetric encoder-decoder splits 8B active parameters for input from 16B for output and shrinks its KV cache 4x to 8x, the same āserve fewer bytes per tokenā logic that lets Sol and Luna undercut their predecessors. Scheduling third: DeepSeekās off-peak rates run at 50% of peak.
Cost per task is the new benchmark
An agent that costs a quarter as much per token but burns three times the tokens, or needs a retry, can still be the expensive choice. That 10-job run is the clearest illustration: Sol finished in two-thirds of the wall-clock time at roughly a third of the spend, and still lost seven of ten tasks on quality.
There is a trap in the other direction. Cheaper tokens invite longer thinking, and reasoning budgets are finite: two max-effort runs on a āpelican on a bicycleā SVG exhausted a 128K reasoning budget before producing output. Efficiency gains get spent, not banked, unless the harness pins effort levels and stop conditions.
FAQ
Is Opus 5.5 better than GPT-6 Sol? On overlapping benchmarks, mostly yes ā 40.0% versus roughly 33% on AutomationBench, plus a higher Artificial Analysis ranking. But no independent test has run both on the same harness at matched effort, so the honest answer is that Opus leads on measured capability while Sol leads on cost per task.
What does a million tokens cost now? Inside OpenAIās lineup: Luna $0.10/$0.50, Sol $2/$10, Astra $10/$50. Add Anthropic at $4/$20 and DeepSeekās V4.1-Flash tier, and frontier-adjacent output pricing spans roughly two orders of magnitude.
Should agent builders switch models this week? Measure first. Instrument cost per completed task, cache hit rate and retry counts on your own workload for a week: the spread between these models on real jobs has been 3x on cost and 7-of-10 on quality preference, wider than any launch table suggests.