Last night’s GPT-6 Sol and Luna note ended with an explicit asterisk: Anthropic shipped Claude Opus 5.5 about ninety minutes earlier, and OpenAI’s cost charts still showed Opus 5 / Fable 5.1. Tonight is that rematch — what you actually route against GPT-6 Astra and Sol once the Claude 5.5 family is on the invoice.
Primary sources: Anthropic launch (Sep 22, 2026), Opus 5.5 overview docs, Digital Applied’s Opus 5.5 vs Astra comparison. Specs and vendor benches are Anthropic’s unless marked independent (Vals AI / Zapier / Digital Applied’s read of both tables).
Spec sheet (named numbers)
| Claude Opus 5.5 | |
|---|---|
| API id | claude-opus-5-5 |
| Family | First Claude 5.5 ship; Sonnet 5.5 / Haiku 5.5 “coming weeks” |
| Context / max output | 1,000,000 / 128,000 |
| Knowledge cutoff | June 2026 |
| Modalities | Text + image input; text out |
| Effort | low → max; default medium; adaptive thinking; thinking always on |
| List price (per 1MTok) | $4 in / $20 out |
| Cache read / write | $0.20 / $5 (5-min write; Digital Applied also lists $8 1-hr write) |
| vs Opus 5 list | Opus 5 was $5 / $25, cache $0.50 / $6.25 |
| Vendor efficiency claim | ~40% lower cost than Opus 5 on typical workloads; output >30% faster; in/out 20% cheaper; cache reads 60% cheaper |
| Fast mode | $8 / $40, up to 2.5× speed (Claude API) |
| Availability | Claude Platform + AWS / Google Cloud / Azure |
| Pre-release external eval | Frontier Design, METR |
| Safeguards (launch shape) | Fable-class cyber / biology / distillation; cyber often re-routes to Opus 4.8; Life Sciences Verification Program; preserved thinking |
Astra stays $10 / $50, and prompts with >272K input bill the whole request at $20 / $75. Opus 5.5 keeps $4 / $20 across its full 1M window — that line alone is why last night’s Sol/Luna post said to re-check Claude before claiming an Anthropic cost win.
What changed for people who run agents
The durable product change is not a new FLOPS sticker. It is list price + token efficiency + safeguard semantics.
Anthropic’s case is Fable-class work cheaper than Opus 5: fewer tokens per task at a lower rate card, with clearer writing that early testers said cut rework. Mid-conversation effort changes can keep the prompt cache (per-message effort beta) — the same agent-shaped cache question as Sol/Luna, different SDK: cost per completed task.
The unattended-loop asterisk: when a cyber or biology safeguard fires, the request falls back to another Claude model and continues. Digital Applied contrasts that with Astra’s API, where monitoring can stop the task. Fallback keeps work moving and changes which model wrote the diff; a hard stop audits cleanly and breaks the run. Plan for the behavior you chose.
Vendor benches (Anthropic’s table)
Anthropic’s published comparison (adaptive thinking at max effort unless noted; Terminal-Bench 4.0 at xhigh for Opus 5.5):
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 Main | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | — | 41.7% |
| GDPval-AA v2.1 (Elo) | 1846 | 1735 | 1708 | 1542 | 1588 |
| AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| HLE with tools | 67.7% | 65.6% | 63.6% | 57.2% | — |
| Terminal-Bench-Science 0.1 | 58.7% | 52.6% | 29.0% | 64.6% | 22.4% |
| OSWorld 2.0 partial | 81.8% | 80.7% | 74.0% | — | — |
Anthropic’s own cost-efficiency claims on that table: at default effort, Opus 5.5 beats Astra on FrontierCode at roughly 20% of the cost per task and matches Astra on Terminal-Bench 4.0 for about 40% of the cost. Treat those as vendor $/task until you reproduce them.
Fine print that belongs in a routing ticket:
- AutomationBench was run by Zapier without fallback models, so safeguard interventions counted as failures — Anthropic says that lowers the score versus production.
- On Terminal-Bench / science benches, cybersecurity and biology tasks that hit safeguards completed on Opus 4.8 / Opus 5 respectively in Anthropic’s setup — that can cut Opus 5.5’s printed score.
- Standard errors are material: Terminal-Bench ±2.6 pts for Opus 5.5; Terminal-Bench-Science ±3.5–5 pts.
- Digital Applied’s cross-read: Opus leads Terminal-Bench / GDPval / HLE by margins larger than harness noise; Astra leads AutomationBench and Terminal-Bench-Science; FrontierCode is within noise. Vendor harnesses disagree by up to 5.4 points on the same Claude model (Opus 5 on FrontierCode: Anthropic 48.0% vs OpenAI 53.4%). Gaps under ~five points are unresolved until your harness says otherwise.
Independent read: Vals AI (launch day)
Digital Applied summarizes Vals AI’s Sep 22 runs:
| Vals benchmark | Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Vals Index (broad) | 66.16% (#4) | 66.61% (#3) |
| Vals RSI Index (24h research) | 37.13% (#1) | 29.03% (#4) |
| ProgramBench fully resolved (of 200) | 37 | 11 |
| Terminal-Bench 2.1 | 87.64% (#1) | 87.27% |
Nearly level on the broad index; Opus’s clearest independent leads are the long tasks. RSI ran each model in its vendor coding agent (Claude Code vs Codex). Vals also used Opus 5 / Opus 4.8 fallbacks for refusals — on SRE Bench, 83% of tasks used them, so treat that slice carefully.
Where it sits next to Astra, Sol/Luna, and the hardware notes
| Purchase | Product | What it buys |
|---|---|---|
| Flagship reasoning (OpenAI) | GPT-6 Astra | Highest OpenAI ceilings; $10/$50; >272K surcharge |
| Default coding agents (OpenAI cost ladder) | GPT-6 Sol | Same methods at $2/$10; cache-steerable effort |
| Volume / extract | GPT-6 Luna | $0.10/$0.50 |
| Flagship / agentic coding (Anthropic) | Claude Opus 5.5 | Fable-class work at $4/$20; $0.20 cache reads; fallback safeguards |
| Prior Claude default | Claude Fable 5.1 | Still the named prior default on this site until you re-bake |
| GPU island math | Vera Rubin NVL72 MLPerf | What the HBM pod does when you do buy the big box |
| Captive accelerator island | Zhenwu V900 | Memory + chip-to-chip + house fabric for someone else’s cloud |
Opus 5.5 does not move KV blocks and does not invent a fabric. It argues that most long coding / knowledge-work agent loops should not burn Astra tokens, and that Claude’s invoice is finally competitive with Sol on the tasks you actually care about — if your eval agrees.
Caveats I would write into any Opus 5.5 ticket
- Reproduce Terminal-Bench / FrontierCode / your repo harness at fixed effort before you price a SKU — vendor $/task uses Anthropic’s research environment.
- Plan for fallback. Cyber/biology safeguards can silently change which model finished the turn; log model id per step.
- Astra still leads some lanes (AutomationBench, Terminal-Bench-Science on vendor tables). Do not paste “Opus wins everywhere.”
- Harness noise. Treat sub-~5 point gaps as unresolved (Opus 5 FrontierCode: Anthropic 48.0% vs OpenAI 53.4%).
- Vals SRE fallback rate. Runs that route refusals to older Opus models are not pure Opus 5.5 scores.
- Fast mode is 2× price. Latency tool, not the default fleet rate card.
- Sonnet/Haiku 5.5 are not out yet. Do not assume the whole ladder prices like Opus.
How I would evaluate the claim if I ran agent fleets
Pick twenty real tasks (migrations, overnight unattended jobs, one citation-checked knowledge pack). Run Opus 5.5, Astra, and Sol at medium and max. Log wall time, tokens, dollars per successful merge, fallback events, and rework minutes. Keep Fable 5.1 on a shadow route for a week. Promote Opus 5.5 only when cost-per-success and rework both move.
Bottom line
On September 22, 2026, Anthropic shipped Claude Opus 5.5 (claude-opus-5-5) at $4 / $20 with $0.20 cache reads, claiming Fable-class work at roughly 40% less cost than Opus 5. The durable facts for a systems reader are the named rate card (including no Astra-style >272K whole-request surcharge), the vendor leads on Terminal-Bench / GDPval / HLE, Astra’s remaining leads on AutomationBench and Terminal-Bench-Science, Vals AI’s nearly-level broad index with long-task edges, and the fallback safeguard contract you must design for.
That is enough to re-open the routing ticket last night’s Sol/Luna post left open. It is not enough to freeze a fleet default without your own twenty-task bake-off.
Create-only. Specs and dates from Anthropic’s Claude Opus 5.5 launch, the Opus 5.5 overview docs, and Digital Applied’s Opus 5.5 vs Astra comparison (including the Vals AI Sep 22 read). Update when a named harness reproduces Terminal-Bench / FrontierCode / AutomationBench against Astra and Sol, or when Sonnet 5.5 / Haiku 5.5 ship.
Sources: Anthropic · Opus 5.5 docs · Digital Applied.
Anpoo Sivanadi, Staff Software Engineer