Claude Opus 5.5 is a meaningful upgrade for long coding and knowledge-work sessions, with a stronger efficiency story than its headline benchmarks. Anthropic lowered the API price to $4 per million input tokens and $20 per million output tokens, made medium the default effort, and says typical tasks cost about 40% less than Opus 5. Artificial Analysis put it at the top of its Intelligence Index; METR called the change over Fable 5.1 incremental.

Start difficult work at medium effort, move to high when needed, and reserve xhigh or max for a measured quality gain. It is worth testing as a premium daily driver, but “best model in the world” is broader than the evidence supports.

Quick facts

SpecClaude Opus 5.5
ReleaseSeptember 22, 2026
API model IDclaude-opus-5-5
Context window1M tokens
Maximum output128K tokens; up to 300K for the Batch API beta header
Standard API price$4 input / $20 output per 1M tokens
Prompt cache$5 per 1M five-minute writes, $8 per 1M one-hour writes, $0.20 per 1M reads
Thinking and effortAdaptive thinking is always on; low, medium, high, xhigh, and max
Default effortmedium
Reliable knowledge cutoffJune 2026
AvailabilityClaude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS
AIHackers-owned CARNot run

Anthropic’s model documentation is the authority for the ID, limits, platform routes, and current prices. A platform listing establishes a route; it does not establish your account entitlement, regional availability, quota, or a free API allowance.

What the launch benchmarks show

Anthropic reports Opus 5.5 ahead of Opus 5 and Fable 5.1 on most rows in its launch table:

EvaluationOpus 5.5Fable 5.1Opus 5GPT-6 AstraGPT-5.6 Sol
Terminal-Bench 4.066.4%55.8%52.3%57.9%37.3%
FrontierCode v1.1 (Main)54.4%50.3%48.0%53.3%47.5%
CursorBench 4.057.8%51.8%46.6%—41.7%
GDPval-AA v2.11846 Elo1735170815421588
AutomationBench40.0%31.4%26.9%41.4%28.8%
Humanity’s Last Exam, with tools67.7%65.6%63.6%57.2%—
Terminal-Bench-Science 0.158.7%52.6%29.0%64.6%22.4%

These are vendor-reported launch comparisons, even where Anthropic used an external benchmark. Unless footnoted, Opus 5.5 uses adaptive thinking at maximum effort. Terminal-Bench uses xhigh for Opus 5.5 and OpenAI’s high-effort Astra result. AutomationBench combines an early-access Opus 5.5 run with public leaderboard rows. Anthropic reports standard errors of roughly 2.6 points on Terminal-Bench and 3.5–5 on Terminal-Bench-Science, so small gaps are not a reliable ranking.

The table still shows a useful pattern. Opus 5.5 is especially strong on agentic coding, professional work, and multi-step research. It is not uniformly first: Astra scores higher on AutomationBench and Terminal-Bench-Science. The rows also mix effort levels and harnesses, and the highest launch setting is expensive enough that the default medium result is the one most buyers should measure.

Independent checks make the claim more useful

Artificial Analysis gave Opus 5.5 a 58 on its Intelligence Index, the highest score it had measured at the time. Its model page identifies the current composite as Intelligence Index v4.3.2, a ten-evaluation set. Opus 5.5 led six of ten evaluations, including Humanity’s Last Exam at 61.4%, SciCode at 66.9%, GDPval-AA at 1846 Elo, and AutomationBench-AA. On its independent Terminal-Bench 4.0 run, Opus 5.5 at max with default fallback scored 59.6%, level with Astra at xhigh. This is a different configuration from Anthropic’s launch row. It remained behind on CritPt, AA-LCR, and GDP.pdf.

The efficiency result is unusually broad. Artificial Analysis says four of five effort levels—medium, high, xhigh, and max—sit on its Intelligence-versus-cost-per-task Pareto frontier. That makes Opus 5.5 easier to route than a model that is attractive only at its most expensive setting. It is still a benchmark frontier, not a promise that every account or workload gets the same cost curve.

That report also complicates “efficient.” At maximum effort, Artificial Analysis measured around 119,000 output tokens per Intelligence Index task for Opus 5.5, versus about 73,000 for Opus 5, 78,000 for Fable 5.1, and 27,000 for Astra. Despite that 1.6x output volume, it found Opus 5.5 max roughly level with Opus 5 on cost per task because the lower rates offset volume. Tool loops, thinking, cache hits, and effort still decide the bill.

Sonar’s pre-release evaluation is a useful code-specific check. On its Java benchmark, Opus 5.5 passed 87.68% of 544 test-backed HumanEval and MBPP tasks, versus 88.6% for Opus 5. It generated 27.5% less code and 40% fewer output tokens, while total findings fell. Bug density rose 12%, concurrency findings 44%, and reasoning tokens were not captured comparably. Sonar plans a post-GA refresh.

METR’s predeployment assessment tested five difficult tasks over ten business days and judged Opus 5.5 a modest improvement over Fable 5.1, with slightly higher expected AI research acceleration but no likely end-to-end automation. Anthropic could review and edit the summary under an unpaid agreement, so treat it as an informed external assessment with a disclosed independence limitation.

The independent picture is therefore strong but bounded: Opus 5.5 belongs in the top tier for coding and knowledge work, its efficiency gain is credible, and its rank depends on benchmark, harness, effort, and cost accounting. AIHackers has not run a controlled repository evaluation or accepted-result study, so CAR remains not run.

Price and effort change the routing decision

The standard input and output rates are 20% below Opus 5. Cache reads are 60% lower, which matters for long-running agents that repeatedly reuse a large prompt. Anthropic also offers a fast mode at $8 input / $40 output per million tokens, with up to 2.5x the speed. Fast mode is a separate research-preview service tier, so do not fold it into the standard cost comparison.

The model’s thinking behavior changed during migration. Adaptive thinking cannot be disabled, and the default effort is medium rather than Opus 5’s high. Anthropic’s prompting guide recommends setting effort explicitly and measuring it. An explicit high setting is a deliberate quality/cost choice, not a migration error. Rebaseline billed thinking tokens and max_tokens when migrating: thinking counts toward the output limit even when its text is not shown.

Use this sequence:

  1. Start routine premium work at medium.
  2. Use high for difficult coding, review, and research when the acceptance bar requires it.
  3. Try xhigh or max only on tasks where your own evaluation shows fewer failures or less human rework.
  4. Track cost per accepted result, tool failures, latency, and review time. A cheaper token is not a cheaper result if it needs another model or a human to repair it.

Why the internet reaction is so positive

Early user reaction clusters around two visible changes: faster interactive tool use and prose that is easier to scan. Reddit users describe fewer detours and better instruction following. These are account-level observations, not prevalence studies, but they line up with Anthropic’s stated communication goal.

One Reddit user published a small 23-task harness with automatic graders, SQL checks, hidden unit tests, and one to three runs per model. The author estimated 100% pass rates for four models, with Opus 5.5 at about $0.09 per passed task and 27 seconds average. Small tiers and API list rates make this a community signal rather than a general benchmark.

The counter-reaction asks whether launch performance will persist and whether max-effort comparisons reflect ordinary subscriber settings. A proposed month-long fixed-prompt benchmark has no result yet; stochastic models require repeated trials and statistical analysis. Keep degradation as an open question.

You can start with Claude Pro

Opus is not only a Max-plan decision. Anthropic’s Pro help lists $20/month in the US and includes Claude Code; current model configuration defaults Pro to Opus 5.5. Local pricing, taxes, usage limits and the installed client still matter. Pro excludes Console API charges.

For focused repository work or knowledge tasks, Pro is a credible tier to try before upgrading. Measure completed work and how often the actual cap blocks it. The September value shortlist compares this route with MiMo, DeepSeek and Luna; the billing checklist helps compare web and app totals.

Reset and cloud credit: useful, but read the terms

With the launch, Anthropic says paid subscription users receive a rate-limit reset that can be saved and used when they choose. The Help Center says it restores the five-hour or weekly limit shown, leaves the normal schedule unchanged, cannot be undone, and must be applied from Claude web or Desktop. Mobile and terminal Claude Code do not show the button.

The official ClaudeDevs update on X specifies Pro, Max, and Team users and says they get a reset to use anytime. The public post does not provide a universal expiry matrix, so check the date and scope shown in your own Settings → Usage before spending it. This is a temporary usage entitlement, not a permanent plan increase.

Anthropic also moved Claude Code cloud sessions out of research preview. The official ClaudeDevs announcement says existing Pro subscribers receive a one-time $100 credit and existing Max subscribers $250, spent first on cloud sessions. It is tied to GitHub-connected cloud work, not general API credit. Account terms control claim and expiry. See September 2026 AI launch offers; this is not a new-subscriber coupon.

One account report displayed $34.38 for a 58-minute cloud task with 109.8 million cached-token reads. That does not establish a billing formula, but it shows why a cloud credit is not $100 of ordinary subscription capacity. Follow the displayed balance and terms before connecting a repository.

Who should use Opus 5.5?

WorkloadStarting pointDecision rule
Long repository migrations, difficult debugging, agentic codingOpus 5.5 mediumKeep it when fewer turns and less review work offset the premium price
Consequential code review or complex researchOpus 5.5 highRaise effort only when your acceptance tests show a gain
Hard terminal tasks with enough budgetOpus 5.5 xhigh or maxMeasure the extra quality against the much higher token use
Routine edits and high-volume subagent workSonnet 5 or a tested value modelEscalate to Opus when the first pass misses behavior or evidence
API migration from Opus 5Pin claude-opus-5-5 and set effort explicitlyTest thinking, tool schemas, cache behavior, and max_tokens before production

Opus 5.5 is a strong candidate for the default premium Claude route. The strongest reason to evaluate it is the combination of leading independent results and lower listed rates. Token savings vary by task and effort. The evidence does not justify replacing every model, setting every task to max, or treating subscription promotions as lasting economics. Run it against accepted tasks, keep a cheaper routine lane, and let the results decide how much of the hype survives.

Sources

Vendor and provider documentation

Independent evaluations

Community reports and discussion

Last reviewed: September 27, 2026 (Asia/Singapore). Pricing, model IDs, benchmark dashboards, plan allowances, promotional eligibility, and provider availability can change independently. AIHackers has not run a site-owned benchmark or accepted-result evaluation for Opus 5.5.