Anthropic just released Claude Opus 5, a model it calls its new flagship. The company is positioning it as a direct answer to a problem that has been building all summer: the best AI is too expensive for most teams to use at scale. Fable 5, Anthropic’s most capable reasoning model, can cost $50 per million output tokens. That price has limited who can afford to run it for serious work, especially as OpenAI and Chinese labs keep slashing rates.
Opus 5 is not a Fable 5 replacement. Instead, Anthropic calls it the best value model it has ever shipped. The pitch is simple: near-Fable 5 smarts for roughly half the token cost. For anyone who needs high-end coding, research, or agentic workflows but can’t swallow Fable 5’s monthly bill, Opus 5 is the intended fix.
Fable 5 Was Powerful, But Too Expensive
To understand Opus 5, look first at what it responds to. Fable 5 launched with record-high reasoning scores and an equally high price tag. The model charges $10 per million input tokens and $50 per million output tokens (The Decoder). For a single complex agent task that burns 100,000 tokens, a user could be paying $5. Even at Anthropic’s own reduced paid tier, that adds up fast.
Meanwhile, OpenAI’s GPT-5.6 Sol entered the market with lower rates and similar performance on several fronts. Anthropic had extended free Fable 5 access to keep subscribers, but the pricing war was clearly pushing on the company’s highest-end model. Chinese competitors like Kimi K3 also started matching or beating certain benchmarks at a fraction of the cost.
Opus 5 arrives as an internal compromise. It uses the same 1 million-token context window and the same token rates as its predecessor, Opus 4.8: $5 per million input tokens and $25 per million output tokens. A new Fast Mode option speeds up responses by 2.5x but doubles the price. The key difference is performance per dollar. According to Matthias Bastian, who analysed early test results for The Decoder, Opus 5 scores higher than Fable 5 on several coding and knowledge benchmarks while staying firmly in the cheaper pricing band.
Opus 5: Same AI Smarts, Half the Token Cost
Independent evaluators confirm the value claim. Artificial Analysis, which worked with Anthropic before the public launch, placed Claude Opus 5 at the top of its Intelligence Index with a score of 61. That edges out Claude Fable 5 (60), GPT-5.6 Sol (59), Kimi K3 (57), and Opus 4.8 (56). The index combines nine tests covering knowledge work, coding, scientific reasoning, and factual accuracy (The Decoder).
Coding is where Opus 5 really lands its punches. On the Artificial Analysis Coding Agent Index, the model at its “xhigh” reasoning tier paired with Claude Code shares the number one spot with GPT-5.6 Sol. Both hit 67. On Terminal-Bench v2.1, which tests AI agents as autonomous engineers in real terminal environments, Opus 5 at “max” effort scored 89 percent, matching the previous leader GPT-5.6 Sol.
Anthropic also gave Opus 5 the ability to check and improve its own work through iteration. In practice, the model can build its own tools in code when the task requires them. That area, agentic coding, is where Opus 5 stands out most clearly. On the Epoch Capability Index for software engineering (SWE-ECI), it ties Fable 5 at 161, outperforming both GPT-5.6 Terra and Opus 4.8. Only GPT-5.6 Sol scores higher in that category.
The knowledge gains are just as clear. On ARC-AGI-3, a test of novel problem-solving, Opus 5 scored 30.2 percent. That is nearly four times the score of GPT-5.6 Sol. On Humanity’s Last Exam, a brutal multi-discipline knowledge check, it ties Fable 5 at 53 percent.
The Max Effort Paradox: When More Compute Hurts
Anthropic offers five reasoning effort settings for Opus 5: low, medium, high, xhigh, and max. The company recommends broad use of the lower settings, claiming they beat earlier Opus models at the same tier while using fewer tokens. For coding and agentic tasks, they suggest starting with xhigh.
But the max effort setting creates a puzzle. On Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index, Opus 5 actually scores slightly worse at max than at xhigh, despite costing more. Cheng-Yuan Lee of FrontierCode explained why. The benchmark judges not just code correctness but whether a model makes only the minimum necessary changes. At higher effort levels, Opus 5 tends to refactor surrounding code even when a tiny fix would do. In one example, the low setting changed exactly one character; xhigh rewrote the whole surrounding module. Those extra changes count as errors in the benchmark, dragging down the score.
This is not a bug. It reflects how deeper reasoning can sometimes over-engineer solutions, a genuine trade-off that teams need to understand before they set the effort dial to max and expect strictly better results.
Knowledge Gains Come With a 50% Hallucination Rate
For all its strengths, Opus 5 has a factual accuracy problem. On the AA-Omniscience benchmark, which tests the correctness of model knowledge claims, Opus 5 improved by 7 points over Opus 4.8 but still trails Fable 5. More striking is the hallucination rate. Because Opus 5 answers more often even when uncertain, its hallucination rate climbs to 50 percent, a 14-point jump from the previous model.
That figure matters for anyone building retrieval-augmented generation or research tools that depend on the model not making things up. While the model’s knowledge breadth has grown, its willingness to answer confidently on weak footing remains an open risk.
Interestingly, independent analysis from Vals.ai using Vibe Code Bench found that the “high” reasoning tier delivered the best real-world coding results. The two highest tiers (“xhigh” and “max”) produced more complex solutions that had more errors overall. Simpler code, produced with less compute, met requirements more reliably. This pattern echoes what the hallucination rate suggests: pushing Opus 5 to its limit can add noise, not signal.
The Frontier Race Remains Razor-Thin
Even with Opus 5’s arrival, no single model owns a clear lead. Epoch AI’s overall Epoch Capability Index puts Opus 5 at 159, just two points below Fable 5’s 161. In software engineering alone, they tie. GPT-5.6 Sol leads in both categories. On the broader Intelligence Index from Artificial Analysis, the top five models sit within four points of each other.
That tightness is why Microsoft CEO Satya Nadella recently said AI models are becoming commoditized. What Opus 5 really delivers is not a new height, but better access to the heights already available. For teams that need top-tier coding and knowledge work without the Fable 5 premium, Opus 5 is a practical answer. But for those who need the absolute lowest hallucination rate or task-specific precision, Fable 5 and even some competitors still hold edges. The model wars are not about who builds the one best AI anymore. They are about who gives the right performance at a cost people can actually live with.
Opus 5 and the Commoditization of AI Models
Opus 5 reflects a market where capability is no longer scarce; price and reliability have become the real differentiators. Anthropic is not asking users to pay more for more. It is asking them to pay the same as Opus 4.8 for something that often works better than its most expensive model. That is a bet that the industry is shifting from chasing the highest benchmark number to delivering the best value for a given task. Whether users bite depends on how much they trust a model that is demonstrably more confident, and sometimes more wrong, than the alternative.
Frequently Asked Questions
How does Claude Opus 5 compare to Fable 5 on price?
Opus 5 costs $5 per million input tokens and $25 per million output tokens, exactly the same as Opus 4.8. Fable 5 costs $10 per million input tokens and $50 per million output tokens. So Opus 5 is roughly half the price per token, and independent tests show it often matches or beats Fable 5 in coding and knowledge benchmarks.
Why does Opus 5 sometimes score worse on the max effort setting?
The max effort setting makes Opus 5 reason more deeply, which can lead it to refactor code broadly even when only a minimal fix is needed. On benchmarks like Frontier-Bench, those extra changes count as errors. For simpler coding tasks, the “high” or “xhigh” tiers often produce better results than “max,” with fewer unnecessary alterations and lower token costs.
What is the hallucination rate of Claude Opus 5?
According to Artificial Analysis, Opus 5’s hallucination rate is 50 percent, up from 36 percent in Opus 4.8. The model answers more often even when uncertain, which improves its knowledge coverage but reduces factual reliability. This makes it less suitable for tasks where strict accuracy is required unless combined with retrieval or verification steps.
