Claude Opus 5.5 is not interesting simply because it tops a launch table. It is interesting because Anthropic is trying to turn frontier-level performance into an everyday workhorse: cheaper per token, more efficient per task, faster to respond, and easier to supervise across long agent sessions.
Watch Matthew Berman's Review and Interview
Source and credit: Anthropic went CRAZY (Opus 5.5), published by Matthew Berman on 22 September 2026. The second half is an interview with Thariq, a member of Anthropic's technical staff. Forward Future collected its wider Opus 5.5 tests and review resources here.
Why Opus 5.5 Looks Like the Workhorse Model
Anthropic's positioning is unusual. It says Opus 5.5 performs at Fable 5.1's level on most work, yet is priced below Opus 5 and needs less compute to serve. In the interview, Thariq calls it a strong daily driver and says he used it across an entire personal-site rebuild.
He gives Fable 5.1 a narrower role: discrete planning, code review, security analysis, and tasks where the cost of a wrong answer is high enough that token price matters less. That division is more useful than a single leaderboard rank:
| Model | Operational role | Use it when |
|---|---|---|
| Claude Opus 5.5 | Daily workhorse | Coding agents, migrations, audits, research, and knowledge work |
| Claude Fable 5.1 | High-consequence specialist | Planning, security review, code review, and expensive-to-reverse decisions |
| GPT-6 Astra | Cross-vendor challenger | Scientific work, demanding computer use, and tasks where your tests show an advantage |
What the Benchmark Table Actually Shows
Anthropic reports 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode 1.1, 57.8% on CursorBench 4.0, and 1846 Elo on GDPval-AA 2.1. Those results place Opus 5.5 ahead of Fable 5.1 and Opus 5 in the published table.
GPT-6 Astra remains ahead on two listed tasks: AutomationBench at 41.4% versus 40.0%, and Terminal-Bench-Science at 64.6% versus 58.7%. Several computer-use and chart-reading rows do not include Astra, so they cannot support a direct win over it.
The footnotes matter. Opus results use different effort settings across tests. Safeguard interventions can trigger fallback models, while Zapier's AutomationBench run counted interventions as failures. Anthropic also says benchmark margins at this capability level are becoming less reliable and that the real-world gap between Opus 5.5 and Fable 5.1 is narrower than the headline scores suggest.
Cost per Task Is the Real Pricing Story
Standard API prices are $4 per million input tokens, $20 per million output tokens, $0.20 per million cache reads, and $5 per million cache writes. That is a 20% input-and-output reduction from Opus 5, while cache reads fall by 60%. Anthropic says fewer tokens and tool calls increase the typical workload saving to about 40%.
Matthew's central point is correct: token price is only one component. An agent that takes ten attempts, repeats context, or writes a long explanation can cost more than a nominally expensive model that finishes cleanly.
Track this instead:
Anthropic's launch examples support the efficiency thesis, although they remain vendor-selected evidence. It reports an HAProxy migration finishing in 9.5 hours versus 12 for Fable 5.1 at 51% lower cost, and a merger-analysis workbook and presentation finishing in 63 minutes versus 93 for Opus 5 at half the cost.
Max Effort Is Not Automatically the Best Choice
The quality-versus-cost charts contain a practical lesson that gets lost in the launch score. On FrontierCode, the medium setting can score better than a much more expensive max run. On Terminal-Bench, medium or high sits closer to the useful upper-left corner: high acceptance at a lower task cost.
More thinking can create diminishing returns, longer latency, and occasionally worse benchmark performance. The safest default is:
- Start at medium. Use a representative task with a strict acceptance check.
- Escalate failures to high. Record whether the extra spend fixes the actual failure mode.
- Reserve max for proven cases. Use it only when repeated evaluation shows a material lift.
- Route simple work downward. Classification, extraction, and short edits usually belong on a cheaper model.
Five Useful Things Anthropic Revealed
1. Claude helps build Claude incrementally
Thariq describes recursive self-improvement as many small gains rather than one autonomous model redesigning itself. Claude writes code, helps with research and evaluation, and contributes to the next system as part of an engineering process.
2. Pacing the frontier does not mean freezing releases
His explanation separates unreleased frontier capabilities from models Anthropic believes it can deploy with robust classifiers, fallbacks, and external evaluation. Opus 5.5 is presented as bringing previously expensive capability to a wider audience, not as the unreleased edge that the pacing proposal targets.
3. Better models can expose an outdated harness
The Claude Code team recently removed parts of its system prompt because old instructions can constrain a more capable model. Anthropic also introduced plugin evaluation so teams can check whether skills and plugins still help the newest model instead of forcing obsolete behavior.
4. Concise communication is operational
Shorter summaries are not merely style polish. When a person supervises 10 or 20 agents, every long status report adds context-switching cost. Opus 5.5 is designed to put the result first and follow explicit writing rules more closely.
5. Tool calling is part of the capability gain
Thariq highlights stronger tool use and custom agent harnesses. That matters because production value comes from reliable actions, verification, and handoffs, not just a better answer inside chat.
What Still Needs Work
The interview is unusually useful because it names unfinished areas. Thariq says combined browser and computer use still needs work so the product can express the capability shown in benchmarks. He also points to personalization, memory, and knowing when to challenge the user without becoming intrusive.
Safety is another boundary. Anthropic says Opus 5.5 achieved its strongest automated behavioral-audit result and improved prompt-injection resistance. Because it approaches Mythos 5.1 in biology and cybersecurity, it ships with Fable-like safeguards and verification programs for vetted research and security organizations. Those controls can affect availability and benchmark behavior.
Finally, the video is launch analysis plus an Anthropic interview, not an independent controlled benchmark. Matthew is enthusiastic, and the guest works for the vendor. Use the official results and interview as evidence about positioning and design decisions; use your own repeated trials to make a purchasing decision.
A Practical Routing Policy
| Task | Start with | Escalate when |
|---|---|---|
| Large code migration or audit | Opus 5.5 medium | Acceptance tests fail or architectural judgment is weak |
| Daily coding-agent work | Opus 5.5 medium | High materially reduces retries on your suite |
| Planning and security review | Fable 5.1 | A second model is required for independent review |
| Browser and computer automation | Opus 5.5 and Astra side by side | Choose from completion rate, interventions, and recovery quality |
| Routine extraction or classification | A cheaper model | Confidence is low or the decision carries material risk |
A Five-Run Evaluation Before You Switch
- Pick five recent tasks your current model completed, including at least one failure.
- Freeze the files, tools, permissions, prompt, and definition of done.
- Run Opus 5.5 at medium and high, then compare with your current route.
- Measure completion, elapsed time, tokens, tool calls, retries, and reviewer corrections.
- Adopt it only where its accepted-result cost is lower or the quality gain justifies the difference.
For more examples beyond this interview, use Forward Future's Opus 5.5 test collection and the site's companion four-build blind comparison.
Video Chapters
| Time | Chapter | Time | Chapter |
|---|---|---|---|
| 00:00 | Introducing Opus 5.5 | 02:10 | Performance benchmarks |
| 06:13 | Pricing and efficiency | 06:44 | Speed improvements |
| 06:52 | Cost per task | 07:16 | Quality versus cost |
| 11:55 | Communication and UI | 12:03 | Safety and alignment |
| 13:03 | Model scalability | 16:26 | Independent intelligence index |
| 17:04 | Anthropic interview | 29:50 | Harnesses, skills, and tool use |