Skip to main content
← All news
AI Models

Claude Opus 5 launches at Opus 4.8 pricing, but changes how developers control the model

By AgentRiot Editorial

Anthropic’s new $5/$25-per-million-token Opus model keeps Opus 4.8 pricing, but default thinking, effort controls, caching changes, and Fast mode make it more than a model-ID swap.

Editorial hero for Claude Opus 5: abstract sculptural form with title text
AnthropicClaudeAI ModelsAgentic AIAPI

Anthropic’s new $5/$25-per-million-token model keeps the old Opus rate card, adds a 1M-token context window, and makes adaptive thinking the default. The price is familiar. The operating model is not.

Anthropic has released Claude Opus 5, a new premium model for complex agentic coding and enterprise work. It is available through the Claude API as claude-opus-5, on paid Claude plans, and through Anthropic’s supported cloud channels. The immediate headline is price: $5 per million input tokens and $25 per million output tokens, the same base rates Anthropic lists for Claude Opus 4.8.

That matters because the launch is not asking existing Opus customers to pay more for a new top-tier SKU. Anthropic is trying a different pitch: keep the familiar Opus price, put more capability behind it, and make the model a practical default for difficult work that is too involved for a fast model but does not justify a long-running frontier-agent budget.

The interesting change is not only the benchmark chart. It is the new controls and defaults developers inherit when they swap model IDs.

The premium tier is becoming a daily driver

Anthropic describes Opus 5 as a step-change over Opus 4.8 in deep reasoning, agentic work, long-horizon tasks, and test-time compute scaling. In its product announcement, the company says the model approaches the capability of Claude Fable 5, its more expensive frontier model, at half the token price.

The current pricing table makes that comparison concrete. Fable 5 is listed at $10 per million input tokens and $50 per million output tokens. Opus 5 is listed at $5 and $25. That is a clean two-to-one difference on standard token pricing, although it does not settle the total cost of a real workload. Tool calls, output length, cache behavior, retries, cloud-provider billing, and the model’s chosen reasoning depth can all change the final bill.

For buyers, that distinction is more useful than a generic claim that a new model is “cheaper.” Opus 5 is not a bargain model. It is a premium model whose base API price did not move upward while Anthropic expanded the context window to 1 million tokens and the maximum output to 128,000 tokens.

Anthropic’s own model overview positions Opus 5 for complex agentic coding and enterprise work. Fable 5 remains the company’s higher-capability option for the most demanding long-running agents. The lineup is therefore less about a single winner than a division of labor: Opus 5 for difficult bounded tasks and serious production workflows; Fable 5 when the workload genuinely needs more autonomy over a longer stretch of time.

Thinking is now on unless you turn it off

This is the migration detail that deserves more attention than it will get in launch-week benchmark threads.

On Opus 4.8, developers had to explicitly enable adaptive thinking. On Opus 5, thinking is on by default. The model decides when and how much to think on a turn, while the effort setting becomes the main control over reasoning depth. The available ladder is low, medium, high, xhigh, and max.

That is a meaningful behavior change, not a cosmetic API option. Anthropic warns that max_tokens caps total output, including both reasoning and visible response text. A workload tuned around non-thinking Opus 4.8 calls may need a larger output allowance or a lower effort setting after a migration.

There is also a new hard boundary: disabling thinking is permitted only at high effort or below. Requests that combine disabled thinking with xhigh or max effort return a 400 error. Anthropic calls this a breaking change from Opus 4.8.

The practical lesson is simple. Treat Opus 5 as a new operating profile, not a drop-in alias. Test representative tasks with existing output budgets, latency expectations, and cost alerts before changing a production default.

The new tools are aimed at agent systems, not chat demos

The release includes a few features that make more sense in persistent agent workflows than in a one-shot chat window.

Opus 5 can add or remove tools between turns while preserving the prompt cache. That capability is currently beta and requires Anthropic’s mid-conversation-tool-changes-2026-07-01 header. For an agent that discovers it needs a new capability mid-task, the appeal is obvious: change the tool set without discarding the conversation state or resending a fixed tool list for the entire session.

Anthropic is also introducing a beta fallbacks: "default" mode. Rather than maintaining a model list for every refusal category, developers can ask Anthropic to apply its recommended fallback model server-side. That reduces configuration work, but it also moves more routing policy into Anthropic’s platform. Teams with strict model-governance or audit requirements should evaluate the behavior rather than assuming “default” means “neutral.”

The minimum cacheable prompt length falls from 1,024 tokens on Opus 4.8 to 512 tokens on Opus 5. That does not guarantee lower spend on every request, but it can make prompt caching available to shorter repeated prompts that previously missed the threshold.

Anthropic is also offering a research-preview Fast mode for the API. The company says it runs at around 2.5 times the default speed, but lists it at $10 per million input tokens and $50 per million output tokens, double the standard Opus 5 rate. Fast mode is not currently available through Amazon Bedrock, Google Cloud, or Microsoft Foundry.

The launch benchmarks are vendor-reported evidence, not a final ranking

Anthropic’s launch material contains a broad comparative scorecard. It reports that Opus 5 leads the models shown on Frontier-Bench v0.1, GDPval-AA v2, ARC-AGI-3, BrowseComp, OSWorld 2.0, FrontierCode v1.1, and AutomationBench. Fable 5 retains higher reported results in several long-horizon or specialist areas.

The headline result is Frontier-Bench v0.1: Anthropic reports 43.3% for Opus 5, compared with 33.7% for Fable 5 and 21.1% for Opus 4.8. VentureBeat’s launch coverage also notes the company’s distinction between bounded evaluations, where Opus 5 performs strongly, and longer autonomous work, where Anthropic still positions Fable 5 as the better fit.

Those numbers are useful launch evidence, but they are still vendor-reported results. They are not substitutes for a public independent leaderboard run under a common harness, nor do they tell a buyer whether the model behaves well with that buyer’s codebase, tools, permissions, and review process. Anthropic itself recommends trying both model classes on representative bounded and long-horizon workloads. That is sensible advice for a launch whose real claim is allocation: use the expensive frontier model less often, not never.

What changes for an existing Opus customer

The cleanest case for Opus 5 is a team that already pays for Opus 4.8 and wants more capability without a higher base token price. The model brings a larger context window, a deeper effort ceiling, shorter cache eligibility, and more flexible in-session tool management.

The migration cost is operational. Teams need to revisit max_tokens, thinking configuration, effort limits, error handling for disabled thinking, cache expectations, and any policy that assumes a static tool list. They should also decide whether Fast mode is worth a doubled token rate for latency-sensitive flows.

Anthropic has made the commercial argument simple: price Opus 5 like the prior Opus tier, then make the distinction about how much work the model can carry. The more consequential question will be whether that new default-thinking behavior produces better completed work per dollar in real agent systems, not just better launch charts.

Sources