On July 24, 2026, Anthropic released Claude Opus 5: a model the company says delivers intelligence approaching Claude Fable 5 across most benchmarks at exactly half the token price. It costs $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8, while Fable 5 runs $10/$50. On Anthropic’s own benchmark suite, Opus 5 outscores Fable 5 on 8 of 13 evaluations. For enterprise teams that have been paying Fable 5 rates for reasoning-intensive agentic work, that math changes the calculus immediately.
What launched
Opus 5 is available today on all platforms, with model ID claude-opus-5 on the Claude API, anthropic.claude-opus-5 on AWS Bedrock, and native availability on Google Cloud and Microsoft Foundry. It is the new default model on Claude Max and the strongest model available on Claude Pro. No mandatory data retention applies for general API access, a point Anthropic flagged specifically for enterprise customers with hard zero-retention requirements.
Three capabilities ship alongside the model:
Five-level effort ladder. Opus 5 supports low, medium, high (default), xhigh, and max effort settings with no beta header required. Lowering effort reduces token spend and latency without a sharp quality drop. Anthropic recommends xhigh as a starting point for most agentic and coding workloads.
Mid-conversation tool changes. Developers can now swap which tools Claude can call within a session without invalidating the prompt cache. Previously, tool swaps forced a full context reset in many agentic pipelines, friction that compounded across long-running workflows.
Automatic safety fallbacks. API requests flagged by safety classifiers on Opus 5 (or Fable 5) now route to the best available model rather than returning a blocked error. Production pipelines no longer fail silently when classifiers fire.
The benchmark picture
Anthropic published a full comparison across 13 evaluations. The table below covers the metrics most relevant to enterprise agentic deployments, with independent validation from Artificial Analysis:
| Benchmark | Opus 5 | Fable 5 | Opus 4.8 | GPT-5.6 Sol |
|---|---|---|---|---|
| Frontier-Bench v0.1 (agentic coding) | 43.3% | 33.7% | 18.9% | 37.5% |
| GDPval-AA v2 (knowledge work, Elo) | 1,862 | 1,748 | 1,595 | 1,738 |
| ARC-AGI-3 (novel problem-solving) | 30.2% | not published | 1.5% | 7.9% |
| OSWorld 2.0 (computer use) | 70.5 | 66.1 | 57.1 | 62.7 |
| AutomationBench (business workflows) | 25.8 | 17.4 | 17.0 | 18.2 |
| CursorBench 3.2 (agentic coding) | 70.1 | 70.4 | 62.3 | 67.1 |
| DeepSWE v1.1 (agentic coding) | 68.8% | 69.7% | not published | 72.7% |
| AA Intelligence Index (independent) | 61 | 60 | 56 | 59 |
Sources: Anthropic launch post, Artificial Analysis, Frontierbeat, VentureBeat
The ARC-AGI-3 result warrants attention on its own. ARC-AGI-3 tests novel problem-solving on tasks the model cannot have memorized during training. Opus 5 scores 30.2%, roughly three times the next-best published result (GPT-5.6 Sol at 7.9%). Opus 4.8 scored 1.5% on the same test. Anthropic notes the size of the lead is surprising and may not translate uniformly into everyday production use, but a jump from 1.5% to 30.2% across a single model generation is not a rounding error. It suggests Opus 5 is genuinely better at novel reasoning tasks, not just better at memorized patterns.
On Artificial Analysis’s Intelligence Index, an independent weighted aggregate across agents, coding, general capability, and scientific reasoning, Opus 5 places first at 61, ahead of Fable 5 (60), GPT-5.6 Sol (59), and Kimi K3 (57).
DeepSWE v1.1 is the benchmark where Opus 5 does not lead. GPT-5.6 Sol tops it at 72.7%, followed by Fable 5 (69.7%), then Opus 5 (68.8%). For teams running fully autonomous GitHub workflows where that 4-point gap reproduces in their specific codebase, GPT-5.6 Sol remains the benchmark leader for that narrow task.
Why the pricing structure matters
Fable 5 has historically pushed teams with cost constraints toward Claude Sonnet 5 or Haiku 4.5 for tasks that actually needed Opus-class reasoning. Opus 5 collapses the justification for that compromise. At $5/$25 per million tokens, it sits below Fable 5’s price floor while scoring above it on most production-relevant benchmarks.
The cost-per-task framing Anthropic uses in its launch materials reflects a real structural shift. On OSWorld 2.0 (computer use), Opus 5 exceeds Fable 5’s peak score at approximately one-third of the token cost. On Frontier-Bench, it outperforms Fable 5 at lower cost per task as well. Even on CursorBench 3.2, where Fable 5 edges it by 0.3 percentage points at max effort, Opus 5 reaches that near-identical result at roughly half the token spend.
For enterprises running recurring workflows at scale, the difference between $5 and $10 input tokens compounds across millions of API calls per day. According to a February 2026 analysis cited by VentureBeat, Claude held roughly 40% of the enterprise large language model market by usage as of late 2025, with Claude Code alone approaching $1 billion in annualized revenue. Holding Opus-tier pricing flat while roughly doubling performance on key benchmarks is the mechanism Anthropic is using to widen the funnel of workloads that are economical to automate.
Pricing breakdown
| Dimension | Rate |
|---|---|
| Standard input | $5 per million tokens |
| Standard output | $25 per million tokens |
| Fast mode input (2.5x speed) | $10 per million tokens |
| Fast mode output | $50 per million tokens |
| Prompt cache write (5-min TTL) | $6.25 per million tokens |
| Prompt cache write (1-hr TTL) | $10 per million tokens |
| Cache hit | $0.50 per million tokens |
| Batch API input | $2.50 per million tokens |
| Batch API output | $12.50 per million tokens |
Source: Anthropic, claudefa.st model spec
Batch processing plus cache hits is the cheapest configuration for large repeated-context jobs. The minimum cacheable prompt also dropped from 1,024 tokens to 512, so short system prompts that were previously too small to cache now create cache entries without any code changes.
Fast mode is available on the Claude API and in Claude Code via usage credits. It is not available on AWS Bedrock, Google Cloud, or Microsoft Foundry at launch, and it cannot be combined with the Batch API.
Safety profile
Anthropic reports Opus 5 as its most aligned Opus model. Misaligned behavior on its internal behavioral audit dropped from 2.85 (Opus 4.8) to 2.30. On an internal benchmark for vulnerability exploitation capability, Opus 5 scores zero. Safety classifiers trigger approximately 85% less often than on Fable 5, which reduces the false-positive load that generates support tickets in enterprise deployments.
Biology-related requests previously blocked by Fable 5 now route to Opus 5 by default (when automatic fallbacks are enabled), since Opus 5 carries a comparable safety profile with better performance on organic chemistry and protein sequence tasks. Cybersecurity remains Fable 5’s domain: Anthropic explicitly did not train Opus 5 on cyber tasks, consistent with its approach on Opus 4.8. Per SiliconANGLE’s coverage, Mythos 5, which Anthropic has not made publicly available, remains the model it evaluates for defensive cybersecurity research under Project Glasswing.
What enterprise buyers should do now
Teams running Opus 4.8 should evaluate a migration. Opus 5 is priced identically and outperforms across every benchmark Anthropic published except DeepSWE v1.1. The upgrade path is operationally simple: update the model ID in your API calls, test the effort level defaults for your use case, and remove legacy self-verification prompts (Opus 5 verifies its own work unprompted). The cache minimum dropping to 512 tokens means some system prompts that previously required restructuring for cache efficiency now qualify without changes.
Teams paying Fable 5 rates for non-cybersecurity work should run a comparison. On 8 of 13 benchmarks, Opus 5 scores higher than Fable 5 at half the price. The exceptions are cybersecurity tasks. For knowledge work, agentic coding, computer use, and multi-step business automation, the switching case is strong. The claude-opus-5 launch on amplifilabs.com provides a production-oriented breakdown of where Fable 5 still makes sense.
The new API features unlock architectures that were previously impractical. Mid-conversation tool changes eliminate the full context reset cost that made dynamic tool-routing architectures expensive at scale. Automatic fallbacks reduce pipeline failure rates from safety classifier errors. Both features are immediately available and enabled in the API call with no additional opt-in.
The effort ladder is the new cost control lever. Dropping from xhigh to high on tasks that do not require maximum reasoning reduces token spend without the quality penalty that came with switching model tiers entirely. Profile your actual workflows against the effort settings before making infrastructure decisions based on list pricing alone. On AutomationBench, Opus 5 at its lowest effort setting still outperforms every prior model on the leaderboard.
The Anthropic enterprise revenue trajectory reported earlier in July established the commercial context: Anthropic’s enterprise ARR has scaled dramatically alongside Claude’s market share gains. Opus 5’s pricing strategy, matching Opus 4.8’s rates while substantially improving performance, is the mechanism for sustaining that trajectory by making more workloads economically viable to automate at Opus-class reasoning quality.
If you want to map Opus 5’s effort tiers and pricing to your specific agentic workload stack, book a call with Enera.