Opus 5.5 and GPT-6 Sol cut the cost of running agents

Two price cuts landed on agent builders inside one week. OpenAI shipped GPT-6 Sol and Luna on September 16 with API prices at half their GPT-5.6 promotional levels, and Anthropic followed on September 22 with Claude Opus 5.5, whose cache reads cost $0.20 per million tokens, 60% below Opus 5. For agents, the cache line decides most of the bill.

Where does an agent's token bill go?#

Every turn of an agent loop sends the same growing context back to the model: the system prompt, the tool schemas, the transcript so far, the pages a fetch tool has read. Only the newest tokens are fresh input. Everything else is a cached read, priced at a fraction of the input rate because the provider still holds it in memory. Anthropic states the consequence in the announcement itself: cache reads "make up the majority of agentic and coding work costs" (Anthropic).

List prices make the arithmetic concrete. Take a session whose stable context reaches 100,000 tokens and then runs 40 turns. That context is paid for 39 more times as cached reads: 39 times $0.05 on Opus 5 is $1.95, and 39 times $0.02 on Opus 5.5 is $0.78, ignoring the single cache write and the output tokens. The session is the same; the largest line on the bill drops by 60%. This is the arithmetic behind both announcements, and it is why an agent that fetches pages, say through searxng-mcp-server, should care more about the cache row than about the headline input price.

OpenAI moved the same row and then built tooling around it. Cached input reads now get "discounts of 90%" (OpenAI), a Prompt Caching Dashboard shows why a request missed the cache, and changing the reasoning effort or the tool list no longer breaks the cache. GitHub reports in the post that these improvements cut the share of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models.

What each vendor cut, and when#

The dates matter to anyone tracking who moved first. OpenAI announced Sol and Luna on September 16 and put both at half their GPT-5.6 promotional prices. Anthropic announced Opus 5.5 six days later, on September 22, with input and output 20% below Opus 5 and cache writes down from $6.25 to $5 per million. Output also arrives more than 30% faster, and a fast mode at up to 2.5x speed is priced at $8 per million input and $40 per million output in Claude Code and on the Claude Platform.

List prices, per million tokens:

The API name is claude-opus-5-5, served on the Claude Platform plus AWS, Google Cloud and Azure (OfficeChai); OpenAI serves gpt-6-sol and gpt-6-luna, both in ChatGPT Work and Codex, with Luna also reaching Free and Go users in the desktop app. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks (TechCrunch). Anthropic also raised the five-hour limits on Pro, Max, Team and seat-based Enterprise plans and added a rate limit reset that subscribers can save and spend when they choose.

Why cost per task moved more than price per token#

Both vendors now publish cost per completed task, and it fell further than any token price. Opus 5.5 at its default effort beats GPT-6 Astra on FrontierCode at roughly 20% of the cost per task, per Anthropic's charts, and one early tester audited a 200,000-line codebase in under three hours where Opus 5 needed more than 20 hours and 2.5x the tokens. On OpenAI's side, GPT-6 Sol at max effort scores 56.4% on Agents' Last Exam, above Opus 5's best, at 60% lower cost, and it lands within 1.1 points of Claude Fable 5's 69.9% on DeepSWE v1.1 at about 80% lower cost per task. A lower price per token and fewer tokens per task multiply.

OpenAI's internal numbers explain the urgency. The median researcher in its rollout consumes more than $600 a day in API-priced tokens, and the 90th percentile passes $7,000. At those levels a 50% cut is a staffing argument, which is why the post positions Sol and Luna for "everyday work" while GPT-6 Astra stays on the frontier.

The fine print under both charts#

Every chart in both posts was run by the vendor whose model won it, and both posts admit limits in footnotes. Anthropic evaluated Opus 5.5 with its production safeguards enabled; when they intervened, "cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5", which the post says "likely reduces Claude Opus 5.5's performance on these benchmarks". The same post warns that "benchmark margins have become a less reliable guide to real-world differences". Builders of security tooling should note the reroute: most cybersecurity work sent to Opus 5.5 is answered by Opus 4.8, and a Cyber Verification Program for verified practitioners opens in the coming weeks.

AutomationBench appears in both posts with different protagonists. Anthropic's table, from Zapier runs without fallback models, puts GPT-6 Astra at 41.4% and Opus 5.5 at 40.0%, a lead for OpenAI's model. OpenAI's own AutomationBench section has Sol at xhigh beating Claude Opus 5 at 9% of the cost per task, plus a footnote that the Fable 5.1 comparison understates that model's cost because it omits Opus 5 fallbacks on roughly 40% of tasks. Same benchmark, different winners depending on whose chart you open.

What changes for the agents on this site#

The support agent on this site sits at the cheap end of this market: gpt-5-nano at $0.05 per million input and $0.40 per million output, about $0.0002 per answer at roughly 1,300 input and 270 output tokens (support agent). Moving it to Opus 5.5 would multiply the per-answer cost with no visitor-visible gain, so the interesting levers are elsewhere. Cache discipline is one: the system prompt and the tool schemas are a stable prefix, and the three-tool-round limit already caps the worst case (case study). Latency is the other. Support is a waiting-customer problem, 5 to 8 seconds today, and a 2.5x fast mode on the mid tier is the release feature most relevant to it.

The next eval round has three planned changes: re-run the conversation fixtures against Sonnet 5.5 when it ships, log cost per answered conversation rather than per token, and keep the scripted fallback wired to the budget switch.

A Monday checklist#

Four checks before anyone reroutes production traffic:

• Read the cache dashboard first: if it shows misses, the new prices are subsidizing a broken prefix. A stable system prompt and a stable tool order are the fix, and both vendors now surface the numbers.
• Evaluate on your own tasks: run the fixtures you already have, log cost per completed task, and treat both vendor charts as one input among several. Both posts concede their margins are unreliable at the top.
• Split the tiers: narrow, high-volume work belongs on Luna-class prices, while long agentic loops that finish in fewer tokens are where Opus 5.5-class economics pay off.
• Check the safeguard notes: if the agent does security or biology work, the model actually answering may be a tier below the one in the config, and that difference belongs in the documentation for audits.

The re-run itself takes an afternoon, and the numbers land in the same trace view the agent already writes.

OpenAI was not done for the week, either: GPT-6 Cyber, a security-focused model, previews within days, and I laid out what that means for agent builders.