Reprice your agent stack in one afternoon

Quality comparisons will take time. The afternoon decision does not.

Three lab posts landed inside about 48 hours. SpaceXAI shipped Grok 4.7 into Cursor at $2 / $6 per million input and output tokens. Anthropic released Claude Opus 5.5 and priced it at $4 / $20, with cache reads at $0.20. Anthropic says it performs at Claude Fable 5.1 level on most work and costs about 40% less than Opus 5 on typical workloads at default settings. About an hour later, OpenAI expanded GPT-6 with Sol and Luna and cut their API prices 50% versus GPT-5.6 promotional pricing: Sol $2 / $10, Luna $0.10 / $0.50.

Simon Willison called the stack a price war. Quality comparisons will take time. The afternoon decision does not. If your Claude routes still default to Opus 5, or your OpenAI volume still sits on GPT-5.6 Sol/Luna promo rates, your cost model is already stale.

Agency and product teams bill against agent hours, retainers, and accepted delivery. Token list prices moved. Effort settings and cache hit rates move the bill further. Keep the ceiling models where they earn their seat (Fable 5.1, GPT-6 Astra). Rebuild the default lane this week.

One afternoon, four moves

1. Re-default Claude to Opus 5.5.

On Claude Platform the model id is claude-opus-5-5. It is live on AWS, GCP, and Azure. Point Claude Code, Cowork, and API routes at it. Anthropic raised five-hour limits on Pro, Max, Team, and seat Enterprise, and added a bankable rate-limit reset. Leave Fable 5.1 for work that still needs that ceiling. Log the change in your harness config, not only in chat.

2. Add Sol and Luna to the OpenAI map.

Sol and Luna are live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. Free and Go get Luna on desktop. OpenAI says these models are not yet in regular Chat. API ids: gpt-6-sol, gpt-6-luna. Route Luna for high-volume product features and cheap loops. Route Sol for everyday agent and coding work. Keep Astra when the task needs the uncompromising tier. OpenAI also improved GPT-6 prompt caching (up to 90% off cached input reads). Harnesses that bust the cache throw away a large part of the cut.

3. Optional Cursor A/B with Grok 4.7.

Grok 4.7 is available in Cursor and Grok Build, plus API and routers, at the same $2 / $6 floor as Grok 4.6, with a fast variant at 2× speed and 2× price. SpaceXAI reports CursorBench 4.0 at 46.3% versus 40.4% on 4.6. Run it on your own long agent jobs. Do not swap the house default on vendor benches alone.

4. Log cost per task, then version the harness.

Pick three internal tasks you already know: one migration-sized, one review-sized, one volume loop. For each model route, record dollars, steps, retries, and rework minutes. Tokens alone hide effort and cache effects. Anthropic’s own framing is that Opus 5.5 costs less per token and uses fewer tokens per task. OpenAI’s Sol/Luna story pairs list-price cuts with caching. When the default model moves, version prompts, skills, and tool contracts the same day. Growing the harness beats stuffing more context into a stale scaffold.

What this does to retainers

If last month’s proposal assumed Opus 5 or GPT-5.6 Sol/Luna floors, the margin math shifted on Sep 22. Recalculate:

• floor cost for your default agent seat this week
• cost corridor you will pass through versus absorb
• whether the retainer still buys readiness, or silently buys yesterday’s token floor

Keep the agent seat as the control plane (identity, permissions, approvals, client context). Reprice the meter under it. The Field note on that commercial split is already live: The agent seat is a control plane.

Monday checklist

1. Flip Claude defaults to Opus 5.5. Note the banked rate-limit reset if you are on a seat plan.
2. Add gpt-6-sol and gpt-6-luna routes. Confirm Work/Codex vs Chat for every seat you sell.
3. Optional: one Cursor A/B day on Grok 4.7 for a long job.
4. Run the same three tasks. Write cost-per-task, steps, and rework into the sheet your proposals already cite.
5. Tag harness/skills with a model-default version. Ship that tag with the next client change window.

If you want that reprice wired into a governed seat and Outcome Spec, Start your build or write hello@mail.22atlas.com.

Linked sources

• Anthropic: Claude Opus 5.5 (2026-09-22): https://www.anthropic.com/news/claude-opus-5-5
• OpenAI: Introducing GPT-6 Sol and Luna (2026-09-22): https://openai.com/index/introducing-gpt-6-sol-and-luna/
• SpaceXAI: Introducing Grok 4.7 (2026-09-21): https://x.ai/news/grok-4-7
• Simon Willison: Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war (2026-09-22): https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/
• Related Field note: The agent seat is a control plane: https://22atlas.com/blog/the-agent-seat-is-a-control-plane/