Anthropic has launched Claude Opus 5, a new flagship model that delivers performance approaching the company’s premium Fable 5 tier while charging token rates half the price. The move comes as pricing pressure intensifies across the AI industry, with OpenAI’s GPT-5.6 Sol and Chinese competitors forcing every major lab to rethink how they balance capability against cost. Opus 5 is now the default model on Claude Max and the most capable model available on Claude Pro, positioning it as Anthropic’s primary vehicle for winning both developer trust and enterprise adoption at scale.
The model maintains the same 1 million-token context window and base token rates as its predecessor Opus 4.8: $5 per million input tokens and $25 per million output tokens. A new Fast Mode increases speed by 2.5x but doubles the price, giving users a clear tradeoff between latency and cost. By contrast, Claude Fable 5 charges $10 per million input tokens and $50 per million output tokens, with correspondingly higher cache write and hit rates. The price gap is substantial, and Anthropic is betting that many users will accept slightly lower peak performance in exchange for dramatically lower operating costs.
Token Rates Tell Only Half the Story
Raw token prices can be misleading without accounting for token efficiency, and Anthropic has faced recurring criticism on this front. Opus 4.7 ended up costing 30 to 40 percent more per task than Opus 4.6, even though both models shared identical base rates. A similar pattern emerged with Claude Sonnet 5, where unchanged token prices masked higher per-task costs. The issue is that newer models often expend more tokens to achieve their results, erasing the apparent savings from flat pricing.
Anthropic is addressing this directly with Opus 5 by introducing five configurable effort settings: low, medium, high, xhigh, and max. These settings allow users to trade off performance against token consumption deliberately, rather than having the model decide how much effort to apply on its own. The company says Opus 5 delivers better value than its predecessor at every effort level, meaning users can match or beat Opus 4.8 performance while spending fewer tokens, or push harder when the task demands it.
In its prompting guide for Opus 5, Anthropic recommends making broad use of the low and medium settings for most tasks, claiming they deliver strong results with a fraction of the token use and latency while still outperforming the same settings on earlier Opus models. For coding and agentic tasks, the company still recommends starting with xhigh, which suggests that complex multi-step reasoning remains token-intensive regardless of model improvements.
Why Opus 5 Sometimes Performs Worse at Maximum Effort
One of the more counterintuitive findings from Anthropic’s own benchmarks is that Opus 5 scores slightly worse at the max effort setting than at the second-highest setting on two key evaluations, despite costing more. The drop appears on Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index. This pattern suggests that pushing the model to its highest effort level can actually degrade output quality, perhaps because it over-elaborates, introduces unnecessary complexity, or second-guesses better initial reasoning.
This is a useful finding for developers: more compute does not always mean better results. The optimal effort setting for Opus 5 appears to be xhigh rather than max for coding and general agentic tasks, which means users can achieve top-tier performance while avoiding both the cost and the quality regression associated with the highest setting. Anthropic’s recommendation to start with xhigh for coding tasks aligns with this data.
Opus 5 Sets Records in Agentic Coding and Knowledge Work
According to Anthropic’s internal benchmarks, Opus 5 posts the highest overall score on the Artificial Analysis Coding Agent Index, narrowly ahead of GPT-5.6 Sol. On Frontier-Bench v0.1, which measures agentic terminal coding, Opus 5 achieves 43.3 percent, beating Fable 5 (33.7 percent), GPT-5.6 Sol (34.4 percent), and its predecessor Opus 4.8 (21.1 percent) by wide margins. For knowledge work measured by the GDPval-AA v2 benchmark, Opus 5 leads with an Elo score of 1,861, ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736).
Opus 5 does not win everywhere. On agentic coding via DeepSWE v1.1, GPT-5.6 Sol leads with 72.7 percent, followed by Fable 5 at 69.7 percent and Opus 5 at 68.8 percent. On health tasks and legal benchmarks, Fable 5 and Mythos 5 outperform Opus 5 respectively. The picture that emerges is one of strong but not universal leadership: Opus 5 excels at open-ended terminal coding and knowledge work, while trailing slightly on structured software engineering sweeps and domain-specific evaluations in medicine and law.
The ARC-AGI-3 Result Is a Major Outlier
The most striking number in Anthropic’s benchmark suite is Opus 5’s score of 30.2 percent on ARC-AGI-3, a benchmark designed to measure novel problem-solving ability without relying on memorized patterns. Opus 4.8 managed just 1.5 percent on this test. GPT-5.6 Sol hit 7.8 percent. That represents roughly a fourfold gap over the next-best model, and it is the kind of jump that researchers and investors pay close attention to.
There is no Fable 5 result for ARC-AGI-3 in Anthropic’s published data, and it is unclear whether such a dramatic lead on a single benchmark will translate into proportional real-world advantages. ARC-AGI-3 is designed to test abstraction and reasoning in novel contexts, areas where even the most powerful models have historically struggled. A jump from 1.5 percent to 30.2 percent is unprecedented in this benchmark family and suggests that Opus 5 incorporates architectural or training innovations that go beyond incremental scaling.
What is ARC-AGI-3 and how does Opus 5 compare on it? ARC-AGI-3 is a benchmark that evaluates AI models on their ability to solve novel reasoning problems without relying on memorized patterns or training data shortcuts. Opus 5 scores 30.2 percent on this benchmark, roughly four times higher than GPT-5.6 Sol (7.8 percent) and twenty times higher than Opus 4.8 (1.5 percent). No other publicly available model comes close to this result, making it the standout data point in Anthropic’s evaluation suite.
Cybersecurity Performance Reveals Deliberate Design Choices
Anthropic says it deliberately did not train Opus 5 on cybersecurity tasks, the same approach it took with Opus 4.8. The result is a model that finds vulnerabilities about as well as Mythos 5 but performs much worse when asked to exploit them. On exploitation-specific metrics, the gap between Opus 5 and Mythos 5 is substantial, and this appears to be a policy decision rather than a capability limitation.
The safety implications are significant. Opus 5’s cyber classifiers trigger approximately 85 percent less often than those on Fable 5, which drew heavy criticism for its frequent interventions that sometimes throttled legitimate security research. Fable 5’s aggressive filtering prompted accusations that Anthropic was invisibly blocking rival AI researchers, a controversy the company acknowledged as a wrong tradeoff. Opus 5 walks a middle path: it permits source code vulnerability research while still blocking binary-based vulnerability scanning, penetration testing, and exploit generation. Blocked requests in Claude.ai, Claude Code, and Claude Cowork default to Opus 4.8 as a fallback, the same approach used with Fable 5.
Opus 5 Builds Its Own Tools When Existing Ones Fall Short
Anthropic describes Opus 5 as substantially better at checking its own work and improving through iteration, with several examples that illustrate emergent problem-solving behavior. In one Frontier-Bench task, Opus 5 received a drawing of a machine part and was asked to create a 3D model in FreeCAD. The catch was that the model had no direct way to view the drawing within its available toolset. Opus 5 responded by writing its own computer vision pipeline to extract the geometry from raw pixels, then used that extracted information to reconstruct the complete machine part. No other model solved this task after five attempts.
In another example, Opus 5 worked on a real bug in a popular open-source package manager. According to Anthropic, the model found the root cause and fixed an edge case that the community patch had missed. A competing model fixed only the surface symptom before reporting the bug as resolved, a pattern that underscores the difference between shallow pattern matching and genuine debugging capability.
An engineer at a trading firm reportedly used Opus 5 to build a market data feed for a new exchange in a single session. Previous models could not complete this task even when given detailed plans. These examples point to a model that can navigate unfamiliar toolchains, compensate for missing functionality by generating its own solutions, and handle multi-step engineering tasks that require both planning and execution.
Scientific Research and Life Sciences Gains
Anthropic positions Opus 5 as the most capable generally available model for scientific research, citing gains over Opus 4.8 across all life sciences evaluations. Standout improvements include a 10.2 percentage point gain in organic chemistry on deriving molecular structures from spectroscopy data, and a 7.7 percentage point improvement on protein-related tasks. The model also shows improved ability to generate visual outputs and analyze visual content such as charts and diagrams, which is relevant for scientific communication and data interpretation.
These gains are consistent with broader industry trends toward models that can assist with experimental design, literature synthesis, and hypothesis generation. For researchers who cannot justify the cost of Fable 5 for routine work, Opus 5 offers a compelling middle ground: near-flagship capability at half the token price, with particular strength in the kinds of open-ended reasoning tasks that scientific work demands.
New Beta Features: Mid-Conversation Tool Changes and Automatic Fallbacks
Alongside Opus 5, Anthropic is releasing two beta features for developers. Mid-Conversation Tool Changes on the Claude Platform allow developers to swap available tools during a conversation without invalidating the prompt cache, which means longer sessions can adapt to changing requirements without incurring the cost of recaching. Automatic Fallbacks on the API route blocked requests to a different model automatically, providing a safety net when the primary model returns an error or refuses a request.
Both features address practical pain points for teams building production systems on Claude. The tool change feature is particularly relevant for agentic workflows where the set of available actions evolves over the course of a session. The fallback feature reduces the operational burden of handling model refusals, which have been a recurring source of friction with earlier Claude models.
What Opus 5 Means for the AI Pricing Landscape
The launch of Opus 5 at half the token price of Fable 5 represents a strategic realignment for Anthropic. The company had previously relied on Fable 5 as its flagship reasoning model, but the pricing was difficult to justify for many production workloads. GPT-5.6 Sol from OpenAI and aggressive pricing from Chinese competitors have compressed the market, forcing every major lab to offer competitive per-token economics alongside strong capability.
Anthropic’s approach with Opus 5 is to compete on value rather than on peak benchmark scores. The model does not win every evaluation, but it wins enough, and it does so at a price point that makes sense for volume deployment. The effort settings give developers fine-grained control over the cost-quality tradeoff, which is essential for applications where latency and token budgets are constrained. And the Fast Mode option provides a path to lower latency for applications that need speed more than cost optimization.
For developers and enterprises evaluating Claude Opus 5, the key question is not whether it matches Fable 5 on every benchmark, but whether the combination of capability, token efficiency, and flexible effort settings delivers better end-to-end value for their specific use cases. On coding tasks, agentic workflows, and knowledge work, the evidence suggests that it often will. On specialized domains like medicine, law, and cybersecurity exploitation, other models retain advantages. The smartest deployment strategy for most teams will be to use Opus 5 as the default workhorse model while reserving Fable 5 or Mythos 5 for tasks that demonstrably require their marginal capability premium. With the introduction of Automatic Fallbacks, that kind of tiered deployment becomes significantly easier to implement.
