Claude Opus 5: Fable-Level AI at One-Third the Price

Anthropic launched Claude Opus 5 on July 24, 2026. The new flagship model matches or beats Fable 5 on agentic coding and knowledge work, costs roughly one-third the price ($5 per million input tokens, $25 per million output), and debuted at number 1 on the Artificial Analysis Intelligence Index. The price-intelligence curve for enterprise AI compressed overnight. The math for anyone routing millions of tokens monthly has changed.

Primary source: Anthropic, "Introducing Claude Opus 5", July 24, 2026.

What happened

Claude Opus 5 is available across all Anthropic platforms as of July 24, 2026: the Claude apps, Claude Code, and the API, at the same price as Opus 4.8. Anthropic describes it as delivering Fable 5-level intelligence at roughly one-third the cost. The claim is not marketing. The numbers hold up.

On Frontier-Bench v0.1, Claude Opus 5 beats every other model and more than doubles Opus 4.8 performance, with lower cost per task. On CursorBench 3.2, it lands 0.5 percentage points from Fable 5's peak while delivering that result at half the cost per task. On ARC-AGI-3, the benchmark for solving novel problems, it scored 30.2%. That is three times the runner-up. On GDPval-AA, a new state of the art. On Zapier AutomationBench, Claude Opus 5's approval rate is roughly 1.5 times the next best model at the same cost per task. On OSWorld 2.0, the computer use benchmark, Claude Opus 5 beats every model at any price tier. It outperforms Fable 5's best result at roughly one-third the cost.

One more number. The 2026 International Mathematical Olympiad (IMO 2026): Claude Opus 5 scored 42 out of 42, per Anthropic. The gold medal threshold was 29.

In life sciences, Claude Opus 5 beats Opus 4.8 across all of Anthropic's internal evaluations. Organic chemistry: +10.2 percentage points. Protein function prediction from sequence: +7.7 percentage points.

Anthropic also announced two beta features. Mid-conversation tool changes preserve the prompt cache on the Claude Platform. Automatic API fallbacks route requests blocked by Claude Opus 5's safety classifiers to another model automatically.

On alignment, Claude Opus 5 scored 2.3 on Anthropic's automated behavioral audit of misaligned behavior. It is the lowest score among all recent Anthropic models. On cybersecurity, Claude Opus 5 matches Mythos 5 at identifying vulnerabilities but trails significantly on exploit writing. On Anthropic's OSS-Fuzz benchmark, Claude Opus 5 finds bugs at a rate comparable to Mythos 5 and produces exploits at a materially lower effectiveness rate.

Claude Opus 5 requires no data retention for general access, matching the policy of previous Opus models.

Why it matters

The AI model market has been split into two tiers since Fable 5 set the price ceiling. The top tier charges $15 to $25 per million input tokens and delivers frontier performance. The lower tier charges $2 to $5 and delivers competitive performance with a visible gap on complex tasks. Claude Opus 5 collapses that divide.

At $5 input and $25 output, Claude Opus 5 delivers performance that, across multiple independent benchmarks, matches or beats models that cost three times as much. The cost per task on the Artificial Analysis Intelligence Index places Claude Opus 5 in the most attractive zone of the scatter plot: high intelligence, low cost per unit of work.

The implication for the CFO is direct. An operation spending $100,000 per month on tokens with a frontier model can, by migrating to Claude Opus 5, cut the inference bill by up to two-thirds with no measurable quality loss on coding, knowledge work, and agentic tasks. The math is simple, but it requires CFOs to evaluate models by cost per task, not by technical score.

For the CTO, the calculation is finer. Claude Opus 5 trails Mythos 5 on offensive cybersecurity and lags Fable 5 on sustained biological research tasks. But it wins on the dimensions that account for most production token volume: coding, automation, knowledge work, and computer use.

The precedent matters. Anthropic repeated with Claude Opus 5 the strategy that made Claude Sonnet the most adopted production model: price a high-intelligence model in the mid-tier range and change the buyer's question. The same price compression dynamic is reshaping the token share landscape. The buyer's question is no longer "which model is the smartest." The question is "which model delivers the intelligence my use case requires at the lowest cost." Claude Opus 5 resets the bar for that question.

What changes in practice

The table below maps the concrete changes Claude Opus 5 introduces across seven dimensions critical for anyone running AI in production: price, performance, cost per task, novel reasoning, computer use, safety routing, and cache management. The old regime is on the left, the new regime on the right.

What to do now

The Claude Opus 5 launch changes the cost equation and routing strategy for any team operating frontier models in production. Three decisions the launch triggers immediately, with the reasoning behind each.

1. Recalculate cost per task, not cost per token. The list price of Claude Opus 5 is identical to Opus 4.8, but Claude Opus 5 generates fewer tokens to solve the same task across multiple scenarios. Anthropic reports 26% fewer tokens on average versus Opus 4.8 in legal use, and one customer reported one-seventh the reasoning tokens with less than half the latency. Migrating from Opus 4.8 to Claude Opus 5 keeps the per-token price and reduces the per-task cost. Migrating from a more expensive frontier model reduces both.

2. Test Claude Opus 5 on your internal benchmark before deciding. The public benchmarks show a tie or win for Claude Opus 5 against Fable 5 on coding, knowledge work, and automation. Your use case may land in the range where Fable 5 still leads, or in the range where Claude Opus 5 wins decisively. The only correct answer is to test with your tasks, in your domain, with your definition of success. The CTO at Box reported 11% gains on data analysis and 17% on due diligence with Claude Opus 5 over Opus 4.8. Your number will be different.

3. If Claude Opus 5 passes the test, recalculate your routing model. Teams operating an AI gateway can add Claude Opus 5 to the model pool and configure routing rules that prioritize it for coding, knowledge work, and automation, reserving the more expensive frontier model for tasks where the performance gap justifies the cost. Migration does not require rewriting integration code. Claude Opus 5 exposes the same API as Opus 4.8.

FAQ

Does Claude Opus 5 replace Fable 5?

On coding, knowledge work, automation, and computer use, Claude Opus 5 matches or beats Fable 5 at half the price. On offensive cybersecurity, Mythos 5 stays ahead. On sustained autonomous biology research tasks, Fable 5 still leads. The answer depends on the use case. Test with your tasks before making the switch.

The price of Claude Opus 5 is the same as Opus 4.8. What changed economically?

The per-token price did not change. What changed is the intelligence per token. Claude Opus 5 solves the same tasks as Opus 4.8 with fewer tokens, fewer tool calls, and less time. The cost per task dropped because the model is more efficient, not because the list price went down.

Does Claude Opus 5 require data retention?

No. The policy matches previous Opus models. Claude Opus 5 has no data retention requirements for general access.

Is 30.2% on ARC-AGI-3 relevant for enterprise use?

ARC-AGI-3 measures the ability to solve problems the model has never seen before, without relying on memorization. For enterprise use, a high ARC-AGI-3 score signals that Claude Opus 5 generalizes better in novel situations. That is exactly the kind of situation that appears when an agent hits an edge case in production or when an analyst requests an analysis with no parallel in the training data. It is relevant. Three times the runner-up score is a gap worth watching.

What to expect

Claude Opus 5 arrives three days after a sequence of frontier launches that put six labs above 50 on the Artificial Analysis Intelligence Index. What sets Claude Opus 5 apart from the other launches this fortnight is the dimension it competes on. It does not compete only on intelligence. It competes on the dimension enterprise buyers prioritize: intelligence per dollar.

Anthropic built Claude Opus 5 as the default model on Claude Max and the strongest available model on Claude Pro. It is not a showpiece model. It is a volume model, positioned to be the workhorse for teams running agents, reviewing code, and automating workflows.

For teams already using the Nexforce Router, Claude Opus 5 enters the available model pool and can be tested in parallel with the current model without changing a line of code. The Router's real-time ranking updates Claude Opus 5's position as independent benchmarks evolve, and routing rules can be tuned to direct coding and knowledge work to Claude Opus 5 while keeping the more expensive frontier model for tasks where the performance delta pays for itself. For the CTO recalculating cost per task after July 24, Claude Opus 5 is the first variable to enter the spreadsheet.

References and Further Reading

• Anthropic, "Introducing Claude Opus 5", July 24, 2026. Primary source.
• Artificial Analysis, "Claude Opus 5: the new leader in agentic knowledge work", July 24, 2026.
• Artificial Analysis, "Opus 5: Fable 5 level intelligence at a lower cost per task", July 24, 2026.
• Artificial Analysis Intelligence Index v4.1, reading of July 27, 2026.
• Anthropic, "Claude Opus 5 System Card", July 2026.