Claude Opus 5 in 2026: price, benchmarks and limits
Claude Opus 5 costs $5/$25 per million tokens and has a 1M context window. Here is what improved, what Anthropic measured, and what the system card warns.
Contents
Claude Opus 5 is Anthropic’s July 24, 2026 model for long-running agents, coding and professional work. Its API price is $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8. The context window is 1 million tokens, and the maximum synchronous output is 128,000 tokens. Anthropic says it delivers substantially better task completion than Opus 4.8 and approaches Fable 5 on some agent benchmarks at roughly half the cost per task. The specifications are product facts; the performance claims come from Anthropic’s own tests. They should not be treated as the same kind of evidence (launch announcement, official model specifications).
- $5 / $25
- input / output priceper million tokens, standard API
- 1M
- context windowone million tokens
- 128K
- max synchronous outputofficial model specification
- ~2.5×
- fast-mode speedat twice the base price
The most consequential number may be the one that did not change. If a new model has the same token rate as its predecessor, the business case moves from “What does one call cost?” to “How many attempts does one finished job require?” A coding agent that runs for 40 minutes and then has to restart turns cheap tokens into expensive work. A more reliable model can cost less per completed task even when it produces more tokens.
What actually changed in Claude Opus 5?
Anthropic describes Opus 5 as “thoughtful and proactive,” with a particular focus on agents that stay on a task for a long time. It becomes the default model on Claude Max and the strongest option on Claude Pro. The API identifier is claude-opus-5. Its reliable knowledge cutoff and training-data cutoff are both listed as May 2026, one detail worth remembering when asking it about events after that date.
The headline capacity is familiar. Opus 5, Opus 4.8 and Fable 5 all have 1M-token context windows and 128K maximum synchronous outputs. Opus 5 does not win by giving developers a dramatically larger container. Anthropic is claiming that it uses the container more effectively, stays coherent through more steps and recovers from trouble more often.
| Specification | Claude Opus 5 | Claude Opus 4.8 | Claude Fable 5 |
|---|---|---|---|
| API input / output | $5 / $25 | $5 / $25 | $10 / $50 |
| Context window | 1M | 1M | 1M |
| Max synchronous output | 128K | 128K | 128K |
| Default API effort | High | Medium | High |
| Anthropic’s positioning | Long agents, coding, professional work | Complex agents and coding | Hardest reasoning and coding |
The default effort level is an easy detail to miss. Opus 5 defaults to high effort in the API and Claude Code, while Opus 4.8 defaults to medium. The dollar rate may be identical, but a model that thinks longer can use more tokens on the same prompt. “Same list price” does not guarantee the same invoice.
What does the $5/$25 price really buy?
The base price is straightforward. Prompt-cache writes cost $6.25 per million tokens for a five-minute cache and $10 for a one-hour cache; cache hits cost $0.50. A large, stable prefix—a codebase map, policy manual or repeated tool definition—can therefore be much cheaper after the first request. Small changes to the prefix can invalidate that advantage, making cache design part of the architecture rather than a billing footnote (Anthropic pricing).
A million-token context window is a ceiling, not a target. Filling it with old logs, duplicated files and loosely related documents creates more material for the model to search and more input to pay for. Retrieval and deliberate context selection still matter. The ability to send an entire repository does not mean that sending the entire repository is the most accurate or auditable choice.
Fast mode is the obvious latency trade. Anthropic says it runs at about 2.5 times standard speed and charges twice the base price. That makes the output rate $50 per million tokens, the same listed output price as standard Fable 5. It can make sense when response time affects revenue, incident recovery or an interactive developer loop. It is a weak bargain for overnight batches that nobody is waiting to read.
Source: Anthropic pricing and launch documents, checked July 25, 2026
| Opus 5 standard | 25 USD |
|---|---|
| Opus 5 fast mode | 50 USD |
| Fable 5 standard | 50 USD |
There is another billing wrinkle. Anthropic’s pricing documentation says the tokenizer used by Claude 4.7 and later can count roughly 30% more tokens for the same text than the older tokenizer, with the actual difference depending on the workload. That warning matters when moving from an older generation. It does not explain an Opus 5-versus-4.8 difference because both are in the newer family, but it is a reason to benchmark real invoices rather than multiply an old token count by the new rate.
How strong are the benchmark results?
Anthropic’s largest improvement claim comes from Frontier-Bench v0.1. The company reports that Opus 5 achieved more than twice Opus 4.8’s performance at a lower cost per task. The footnote changes how that sentence should be read: Anthropic ran the evaluation internally with the mini-SWE-agent harness on Google Kubernetes Engine, averaged reward across five attempts, and allowed Opus 4.8 fallbacks when Opus 5 refused a task. That makes it closer to a service-level evaluation than a clean, single-model duel.
On CursorBench 3.2 at maximum effort, Opus 5 finished within 0.5% of Fable 5’s peak at half the cost per task. Anthropic also reports a score three times the next-best model on ARC-AGI 3, roughly 1.5 times the next-best pass rate at the same cost on Zapier AutomationBench, and an OSWorld 2.0 result above Fable 5’s peak at a little over one-third of the cost. Its internal science tests improved by 10.2 percentage points in organic chemistry and 7.7 points on protein variant-function prediction relative to Opus 4.8.
Source: Anthropic internal evaluations, published July 24, 2026
| Organic chemistry | 10.2 pp |
|---|---|
| Protein variant-function | 7.7 pp |
These results are useful directional evidence, not independent replication. Agent benchmarks are unusually sensitive to the harness, tool permissions, stopping rules, retry budget and scoring method. Allowing a fallback model answers a practical question—“Can the service finish the job?”—but not the narrower question “What can Opus 5 alone do?” The customer quotations in the announcement are also early-access reports selected by the vendor, not controlled comparisons.
The right response is not to discard the numbers. It is to preserve their labels. A vendor-run result with disclosed conditions is better evidence than an unexplained superlative, but it is still a starting hypothesis for a production test.
Why could long agent runs feel different?
Small reliability gains compound across long workflows. In a deliberately simplified example, if each of 20 required steps has a 95% chance of succeeding, the chance that all 20 succeed is about 36%. Raise each step to 97%, and the combined result is about 54%. Real steps are not independent and good agents can recover, so those figures are not an Opus 5 forecast. They illustrate why a two-point improvement can be almost invisible in chat and material in a multi-hour job.
Opus 5 also launches with API features aimed at keeping long sessions alive. A beta for mid-conversation tool changes lets a system add or remove tools without restarting the conversation or necessarily throwing away the prompt cache (tool-change documentation). Server-side fallback can route an eligible Opus 5 request to Opus 4.8 when the new model cannot serve it (fallback documentation).
Both features improve resilience and complicate auditing. A team should record the model that produced the final answer, the reason for any fallback, the active tool set and the time of each tool change. Otherwise a dashboard labeled “Opus 5” may quietly include Opus 4.8 outputs. That can matter for reproducibility, regulated decisions and regression analysis.
Higher autonomy also calls for clearer instructions, not vaguer ones. Anthropic’s Opus 5 prompting guide recommends explicit success criteria, checkpoints and boundaries. “Fix the repository” gives a long-running agent a large search space. Naming the allowed files, required tests, forbidden changes and completion evidence gives it a finish line.
Did safety improve without making the model unusable?
Anthropic’s 194-page system card describes Opus 5 as the most aligned model in its automated audit so far. Its aggregate misaligned-behavior score was 2.3, where lower is better. Across roughly 3,200 investigations per model, most measured rates for deception, user manipulation and power-seeking behavior were low. The report says evaluators did not observe sandbagging, malicious action or attempts to evade oversight, and resistance to prompt injection improved.
The model is not unrestricted. Its vulnerability-finding ability came close to Mythos 5, but exploit development remained far behind. Anthropic allows source-code vulnerability discovery while blocking binary scanning, penetration testing and exploit generation in relevant product surfaces. The company expects classifiers to intervene about 85% less often than with Fable 5, but flagged requests in Claude.ai, Claude Code and Cowork fall back to Opus 4.8 by default rather than simply passing through.
Anthropic classifies it at CB-1, not CB-2, for cyber capability and keeps the same ASL-3 protection level used for Opus 4.8 in biology. The system card says it does not cross the automated AI-research threshold, is not close to replacing Anthropic’s research scientists or engineers, and does not meet the threshold for dramatic AI-attributable acceleration. “Long-running agent” should not be casually translated into “autonomous superintelligence.”
What does Anthropic’s launch page leave out?
Three caveats deserve more attention than they receive in the announcement.
First, Anthropic produced the major performance charts and chose the early-access testimonials. There is no reason to assume the results are fabricated, but there is also no basis yet to call them independently reproduced. Frontier-Bench and the two science evaluations are internal runs. Benchmark setup, retry policy and fallback behavior can change the result.
Second, the system card contains an awkward accuracy finding. Opus 5 was more accurate overall, yet it sometimes stated an answer confidently while uncertain and showed slightly more factual hallucination than Opus 4.8. A model that writes longer, more coherent answers can make a false premise more persuasive and more expensive to review. Legal, medical and financial workflows still need fixed sources, calculation checks and human approval.
Third, Opus 5 is not more capable overall than Fable 5. Anthropic’s claim is that it approaches the frontier at a lower cost on selected work, not that it replaces the flagship everywhere. Mythos 5 remains stronger in exploit development and long-running autonomous biology research. Opus 5 is best understood as a much stronger model at the old Opus price point.
| Launch claim | Condition that belongs beside it |
|---|---|
| More than 2× Opus 4.8 | Internal Frontier-Bench run, five-attempt mean, 4.8 fallback on refusals |
| Near Fable 5 | Within 0.5% of its CursorBench 3.2 peak under a specific setup |
| Same $5/$25 price | Default effort is High, so tokens per task can change |
| About 85% fewer classifier interventions | Product restrictions and fallback behavior still apply |
| Most aligned automated audit | Factual hallucination was slightly higher than Opus 4.8 |
Who should use Opus 5 now?
The clearest candidates are teams running long code changes, multi-stage research or internal automations where a late failure is expensive. Existing Opus 4.8 users have a low-friction test because the list price is unchanged. Short classification, simple summaries and high-volume customer support should still be compared with cheaper models. A million-token context window is not a reason to make the premium model the default for every request.
Test on work that has actually failed, not only on polished demos. A useful migration set records first-pass completion, retries, input and output tokens, tool errors, unsupported claims, human correction time and the final model after any fallback. The comparison metric is total cost per accepted task, not cost per request. Run fast mode as a separate treatment only where latency has a measurable value.
This is also part of the wider open-model versus cloud-model trade-off. Opus 5 keeps the closed-API model but tries to improve reliability without raising its tier price. The dependence goes below software too: our analysis of why advanced AI still relies on Korean HBM explains why longer contexts and agent runs turn model progress into a memory and infrastructure problem.
Opus 5 looks like a sensible successor to Opus 4.8 if it finishes more of a team’s real jobs on the first attempt. But granting broader autonomy before measuring hallucinations, tool failures and fallback behavior would reverse the order of operations. Better capability should earn more trust from evidence; it should not receive it automatically from a launch chart.
FAQ
- How much does Claude Opus 5 cost?
- The standard API rate is $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. Fast mode is about 2.5 times faster and costs twice the base rate.
- What is Claude Opus 5's context window?
- It has a 1-million-token context window and a 128,000-token maximum synchronous output. That is a capacity limit, not a reason to fill every request with one million tokens.
- Is Claude Opus 5 better than Fable 5?
- Not across the board. It came within 0.5% of Fable 5's peak on Anthropic's CursorBench 3.2 run at half the cost per task, but Anthropic's system card says Opus 5 is not more capable overall.
- Should Opus 4.8 users migrate immediately?
- Teams running long coding or agent tasks should test it because the token rate is unchanged. Compare first-pass completion, retries, tool errors, hallucinations, and total cost per successful task before switching production traffic.
