Anthropic on Friday released Claude Opus 5, its most capable publicly available AI model to date, at half the price of its immediate predecessor — a deliberate response to mounting enterprise pressure to prove that advanced AI can deliver measurable value without runaway costs.
The model costs $5 per million input tokens and $25 per million output tokens, according to Anthropic, compared to double those figures for Claude Fable 5. Anthropic said Opus 5 outperforms Fable 5 on coding and knowledge work evaluations, and described it as a model “designed to be used every day.”
The Three Facts That Matter
- Performance beats the previous generation — on the benchmarks that businesses care about most. Anthropic said Claude Opus 5 outperforms Fable 5 specifically on coding and knowledge work evaluations. These are the two task categories where enterprises typically see the most direct productivity return, making the comparison commercially significant rather than purely academic.
- Anthropic is deliberately capping the model’s most dangerous capabilities. The company said Opus 5 is not state-of-the-art for “risky, dual-use capabilities” — its term for areas like offensive cybersecurity and biology research. Anthropic said those capabilities remain concentrated in Claude Mythos 5, a model released in June that was restricted from general access. Anthropic head of product management for research Dianne Penn told CNBC the company continues to collaborate with government agencies during pre-deployment testing, including for Opus 5. This is a meaningful distinction: Anthropic is effectively maintaining a two-tier model architecture, separating frontier capability from general-purpose deployment.
- The pricing reflects a fundamental shift in what enterprise customers are demanding. “Enterprises, in our feedback and with our customer base, are looking for value,” Penn told CNBC. “If it’s a cheaper model or a cheaper offering, but it’s not accomplishing a similar level of quality, it’s actually not useful.” The framing is significant: Anthropic is not competing on price alone, but on price-adjusted quality — a more defensible position as competition from Microsoft, Amazon, Google, and Chinese AI startups intensifies. This dynamic is consistent with the broader trend documented across the industry, where corporate AI buyers are moving from hype-driven experimentation to hard-nosed cost discipline.
Taken together, Anthropic’s dual strategy — releasing a cheaper high-performance model for everyday enterprise use while keeping its most sensitive frontier capabilities restricted to Mythos 5 — represents something the source reporting does not directly articulate: the company is effectively operating two parallel product lines with distinct risk profiles. One is calibrated for scale and commercial adoption; the other is calibrated for national-security scrutiny. That bifurcation is likely to become a template competitors are forced to consider, as regulators and enterprise compliance teams increasingly demand to know not just what a model can do, but what it has been deliberately prevented from doing.
How Claude Opus 5 Compares to Competing Flagship Models
Opus 5 enters a market crowded with frontier models from rivals who have also been racing to reduce per-token costs. The table below compares publicly available positioning across the leading general-purpose enterprise models, based on company disclosures and widely reported benchmark results.
| Model | Developer | Key Strength (per company claims) | Notable Restriction / Caveat | General Pricing Position |
|---|---|---|---|---|
| Claude Opus 5 | Anthropic | Coding & knowledge work; designed for daily use | Not state-of-the-art on dual-use/cybersecurity tasks | $5 / $25 per million tokens (input/output) |
| Claude Fable 5 | Anthropic | Previous best general-purpose model; broad capability | Subject to export control disruption in June 2025 | ~2× the cost of Opus 5 (per Anthropic) |
| GPT-4o / o-series | OpenAI | Multimodal breadth; agent frameworks | Cost has been a recurring enterprise complaint | Varies by tier; competitive with frontier pricing |
| Gemini 1.5 Pro / 2.x | Google DeepMind | Long-context window; integration with Google Cloud | Benchmark comparisons contested by third parties | Competitive; Google has cut prices on Flash variants |
Anthropic is also facing pressure from a cohort of lower-cost Chinese AI startups, which the company acknowledged to CNBC, alongside competition from Microsoft and Amazon — both of which distribute Anthropic’s own models through their cloud platforms, creating a complex competitive dynamic in which partners and rivals are often the same entities. The broader geopolitical dimension is worth noting: U.S. officials have raised concerns about Chinese AI models and intellectual-property practices, adding regulatory texture to what might otherwise look like a straightforward pricing war.
Enterprise buyers are also wrestling with a more fundamental concern that goes beyond sticker price. As Microsoft’s Satya Nadella has flagged publicly, companies using third-party AI models face a dual cost: direct token spend and the risk of exposing proprietary data. Anthropic’s “designed to be used every day” positioning for Opus 5 implicitly addresses the first concern; the second remains an open question for any hosted AI offering.
The model’s launch also arrives in the shadow of a turbulent two months for Anthropic. In April, the company released Claude Mythos Preview — a model it said demonstrated exceptional offensive cybersecurity capability, including the ability to discover and exploit software vulnerabilities. That disclosure triggered intense scrutiny. When Anthropic followed up in June with Mythos 5 and the general-purpose Fable 5, both models were briefly disabled within days of release to comply with a government export control directive citing national security authorities. Access was restored after roughly two weeks of negotiations, according to CNBC. The episode underscores why Anthropic’s decision to explicitly describe what Opus 5 cannot do — rather than only what it can — is as much a regulatory communication strategy as a product specification. The risks of capable AI agents operating in sensitive environments are increasingly concrete, as recent incidents involving autonomous AI agents have demonstrated.
The Implications That Matter
- Halving the price of a best-in-class model resets enterprise expectations industry-wide. Competing AI labs will face renewed pressure to justify their own pricing tiers, particularly for models targeting the same coding and knowledge-work use cases where Opus 5 now claims a benchmark advantage.
- Anthropic’s two-tier model architecture — general-purpose versus dual-use frontier — may become a compliance template. By clearly delineating which capabilities are excluded from Opus 5 and reserving them for the more restricted Mythos 5 line, Anthropic is building a paper trail that enterprise compliance teams and government partners can audit, a structural advantage over competitors who have not made similar distinctions explicit.
- The export control episode that preceded this launch is an unresolved institutional risk. Anthropic was forced to disable two models after government intervention; the fact that Opus 5 is being released with explicit government collaboration during pre-deployment testing signals the company is attempting to institutionalize that relationship rather than treat it as an obstacle. Whether that cooperation translates into durable regulatory goodwill — or simply into a more complex approval process for future frontier releases — remains the central variable to watch.
- The “value, not just cheapness” framing from Anthropic’s own product leadership is a strategic bet on quality retention at lower price points. If Opus 5 benchmarks hold under independent scrutiny, Anthropic will have demonstrated that cost reduction and capability advancement are not in tension — a claim that, if validated, reshapes the competitive calculus for every lab currently burning capital to stay at the frontier.
