Claude Opus 5.5: Why the Model Recommendation Just Changed
Opus 5.5 now performs at Claude Fable 5.1's level on most work while costing 40% less than the old Opus 5, and per-token it's now cheaper than Opus 5 was three weeks ago. That changes the advice from this site's last model comparison.
Follows directly from this site's own Fable 5.1 vs Opus 5 vs Sonnet 5 comparison published three weeks earlier, this piece exists because that recommendation changed and the reasons why are worth being precise about.
Claude Opus 5.5, released September 22, 2026, now performs at Claude Fable 5.1's level on most work while costing $4/$20 per million input/output tokens, 20% less than Opus 5's $5/$25 and 40% less than Anthropic's own stated comparison. It also defaults to medium effort instead of high, a change that silently affects code carried over from Opus 5 without an explicit effort setting.
What Changed
Anthropic released Claude Opus 5.5 on September 22, 2026, and replaced Claude Opus 5 as the default recommendation in its own model comparison table: "start with Claude Opus 5.5 for most workloads." Opus 5 is still available but now sits in Anthropic's legacy model list rather than the current lineup.
| Claude Opus 5.5 | Claude Opus 5 (legacy) | Change | |
|---|---|---|---|
| Input per MTok | $4 | $5 | -20% |
| Output per MTok | $20 | $25 | -20% |
| Cache reads per MTok | $0.20 (5% of input) | $0.50 (10% of input) | -60% |
| Cache writes per MTok | $5 | $6.25 | -20% |
| Default effort | Medium | High | Changed |
| Generation speed | Baseline | 30%+ slower | Opus 5.5 faster |
| Retirement (not sooner than) | Sep 22, 2027 | Jul 24, 2027 | n/a |
Anthropic's own framing of the headline change: Opus 5.5 "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5." That's a capability jump and a price cut landing on the same release, not a tradeoff between them.
The Benchmark Numbers
Anthropic's own published benchmark deltas, Opus 5.5 versus Opus 5:
| Benchmark | Opus 5 | Opus 5.5 | Delta |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 52.3% | 66.4% | +14.1 pts |
| GDPval-AA v2.1 (knowledge work, Elo) | 1708 | 1846 | +138 Elo |
| OSWorld 2.0 (computer use) | 74.0% | 81.8% | +7.8 pts |
The Terminal-Bench 4.0 jump is worth putting next to this site's earlier Fable 5.1 vs Opus 5 comparison: Fable 5.1 scored 55.8% on that same benchmark against Fable 5's 42.0%. Opus 5.5's 66.4% now beats both Fable 5.1 and Fable 5 on agentic coding, at a fraction of Fable 5.1's $10/$50 pricing.
This Changes the Recommendation From the Last Piece
The earlier piece's framework was: start with Opus 5 at effort high, move to Fable 5.1 only after Opus 5 falls short on your own evals, and reserve Fable 5.1 for genuinely long-horizon agentic work where its benchmark gains actually concentrated.
That framework still holds structurally, but the numbers underneath it moved. Opus 5.5 now performs at Fable 5.1's level "on most work" per Anthropic's own claim, while costing 60% less than Fable 5.1's $10/$50 pricing. The practical effect: the set of workloads that genuinely need to reach for Fable 5.1 over Opus 5.5 is smaller than the set that needed to reach for it over the older Opus 5, since the baseline they're being compared against got both better and cheaper at the same time.
| Claude Fable 5.1 | Claude Opus 5.5 | Claude Sonnet 5 | Claude Haiku 4.5 | |
|---|---|---|---|---|
| Input / output per MTok | $10 / $50 | $4 / $20 | $2 / $10 | $1 / $5 |
| Context window | 1M tokens | 1M tokens | 1M tokens | 200K tokens |
| Default effort | High | Medium | High | Not supported |
| Thinking | Adaptive, always on | Adaptive, always on | Adaptive | Extended (manual) |
The updated default, in order: start with Opus 5.5, confirm on your own evals before reaching for Fable 5.1, and treat Fable 5.1 as the exception for long-horizon agentic work rather than the aspirational tier Opus models are compared against.
The Operational Gotcha: Default Effort Changed
Every other current-generation model (Fable 5.1, Sonnet 5) defaults to effort high when the parameter is omitted. Opus 5.5 defaults to medium.
Code migrated from claude-opus-5 to claude-opus-5-5 by changing only the model string, with no explicit output_config.effort set, will silently run at a lower effort level than it did before. That's not necessarily worse, medium effort on Opus 5.5 may well outperform high effort on the old Opus 5 given the benchmark deltas above, but it's a behavior change that won't show up as an error, just as a quality or verbosity difference that's easy to misattribute to something else during a migration.
// Migrated model string, but the effort default is now different.const response = await client.messages.create({ model: "claude-opus-5-5", max_tokens: 16000, // No output_config.effort set: this now runs at "medium", not "high". messages: [{ role: "user", content: "..." }],});// Explicit, if you want to match Opus 5's old default behavior:const response2 = await client.messages.create({ model: "claude-opus-5-5", max_tokens: 16000, output_config: { effort: "high" }, messages: [{ role: "user", content: "..." }],});Anyone migrating existing Opus 5 code should set effort explicitly rather than relying on the default carrying over unchanged.
What About the "Good at Explainer Videos" Buzz?
A secondary story circulated alongside the Opus 5.5 launch claiming particular strength at generating explainer videos. Checked directly against Anthropic's own Opus 5.5 announcement: there's no mention of video or explainer-content generation as a new capability. The announcement's stated improvements are agentic coding, knowledge work, and communication clarity, not multimedia generation. That claim traces to a third-party site, not to Anthropic's own materials, and this piece isn't repeating it as a confirmed capability.
Frequently Asked Questions
What's actually new in Claude Opus 5.5 versus Opus 5?
Lower pricing ($4/$20 per million input/output tokens versus $5/$25, a 20% cut, with cache reads down 60%), meaningfully higher benchmark scores (+14.1 points on Terminal-Bench 4.0, +138 Elo on GDPval-AA v2.1, +7.8 points on OSWorld 2.0 computer use), over 30% faster generation, and a changed default effort level (medium instead of high). Anthropic states it now performs at Claude Fable 5.1's level on most work while costing 40% less than Opus 5.
Is Claude Opus 5.5 cheaper than Claude Opus 5?
Yes. Input tokens dropped from $5 to $4 per million (-20%), output from $25 to $20 per million (-20%), and cache reads from $0.50 to $0.20 per million (-60%, since cache reads are now priced at 5% of input cost instead of the standard 10%).
Why did the default effort level change, and does it affect existing code?
Claude Opus 5.5 defaults to effort medium, while Opus 5 defaulted to high. Code migrated to the new model string without explicitly setting output_config.effort will run at a different effort level than it did on Opus 5, silently, with no error. Set effort explicitly during migration rather than relying on the previous default carrying over.
Does Claude Opus 5.5 generate videos?
Not according to Anthropic's own announcement, which makes no mention of video or explainer-content generation as a capability. That claim originated from a third-party site, not from Anthropic's own materials, and isn't confirmed by the primary source.