Claude Opus 4.8, released 28 May 2026: what the launch delivered and where it stands since Opus 5.5

On May 28, 2026, Anthropic released Claude Opus 4.8, at the time by its own account the most capable generally available model the company had built. Unlike the months of rumor-mill speculation around leaked version strings, it was not conjecture but a real product with documented benchmarks, an API identifier and a price tag. It held the top spot for only a few weeks.

Claude Sonnet 5 arrived on 30 June 2026 and Claude Opus 5 on 24 July 2026, the latter positioned well above Opus 4.8 at unchanged Opus pricing (5 dollars input and 25 dollars output per million tokens) with a one-million-token context window. Claude Fable 5.1 followed on 1 September for long-horizon agentic work, and on 22 September Claude Opus 5.5 at $4 input and $20 output per million tokens. On 16 October 2026 the documentation listed Opus 4.8 as legacy; it remains available at its list price of $5/$25, with retirement no sooner than 28 May 2027. For a new integration there is therefore no price argument for 4.8 left, only the effort of re-validating a system that already runs stably on it.

The short version for those in a hurry: Opus 4.8 was not a leap into a new generation, but a precisely placed point release on top of Opus 4.7. The real progress lay not in a single benchmark record, but in a property far more important for the autonomous use of AI: honesty. Anthropic explicitly called honesty one of the most prominent improvements in 4.8, and that has tangible consequences for anyone who sees AI not as a chat toy but as a working part of their processes.

What Opus 4.8 measurably did better

Let us start with what can be quantified. On SWE-bench Pro, at release the toughest benchmark for evaluating AI coding agents, Opus 4.8 reached a record 69.2 percent. For comparison: its predecessor Opus 4.7 scored 64.3 percent, OpenAI's competing GPT-5.5 reached 58.6 percent. That was a lead of more than ten percentage points over the strongest competitor at the time, in a field where every single point is hard-won.

Beyond pure coding there are concrete numbers too, though they come from customers quoted in Anthropic’s announcement. One of them (Miguel Gonzalez, tech lead) measured 84 percent on Online-Mind2Web, a benchmark for browser and computer-use tasks. On Harvey’s Legal Agent Benchmark Opus 4.8 became the first model to break the 10 percent mark under the strict „all-pass" standard, where a test case only counts if every single sub-step was solved correctly. And on another customer’s Super-Agent benchmark (Kay Zhu, co-founder and CTO), Opus 4.8 was the only model to complete every test case end to end, beating GPT-5.5 at parity on cost.

These numbers were impressive, but they were not the reason we considered the release relevant. Benchmark peaks age quickly, as this release itself showed: barely eight weeks later, Opus 5 already sat above it. More interesting is the question of how reliably a model works when no one is watching.

The real leap: honesty as a feature

Anthropic called honesty one of the most prominent improvements in Opus 4.8 and backed that up with a concrete metric: according to internal testing, the model is roughly four times less likely than Opus 4.7 to let errors in its own code pass unremarked. It is also markedly more willing to openly name uncertainties about its own work instead of making unsupported claims. The rate of misaligned behavior is substantially lower than its predecessor's.

Why does this matter more than a benchmark point? Because the most expensive mistake an AI agent makes is not the one it makes — but the one it makes and then conceals. A model that builds a subtle bug into a migration and reports it as „done" costs a company more than a model that honestly says: „I could not reliably verify this part." This is exactly where 4.8 came in. For autonomous use — wherever AI works independently across multiple steps — calibrated self-assessment is the precondition for trust. Anthropic itself described the model as more mature in judgment, more honest about its own progress, and able to work independently for longer than its predecessors.

From our practice: this precise point decides whether an AI workflow is allowed to go into production at a client or not. An honest model can be safeguarded with clear escalation rules („if uncertain, escalate to a human"). An overconfident model undermines every control layer you build around it.

The technical innovations in detail

The API identifier is claude-opus-4-8. The model supports a 1-million-token context window by default on the Claude API, Amazon Bedrock, Vertex AI and Microsoft Foundry (release notes), as well as a maximum output length of 128,000 tokens. It uses the same tools and platform features as Opus 4.7 — so existing code keeps running without modification. New with 4.8 were these building blocks:

• Effort Control. Since 4.8, users on claude.ai and in Cowork can control how much effort the model puts into a task — higher levels allow more thorough reasoning, lower ones save tokens and time. The default value of the effort parameter is high on all surfaces, including the Claude API and Claude Code.
• Adaptive thinking as the only thinking mode. As with 4.7, there are no fixed „thinking budgets" anymore. The model decides per response whether a task requires reasoning: on simple lookups it answers directly, on complex multi-step problems it reasons first. On mixed workloads this noticeably saves thinking tokens compared to 4.7 at the same effort level.
• Fast Mode. Launched as a research preview on the Claude API. With speed: "fast", the same model delivers up to 2.5x more output tokens per second — at a premium surcharge.
• Mid-conversation system messages. Opus 4.8 accepts role: "system" messages in the middle of a conversation, directly after a user turn. You can therefore update instructions later without repeating the entire system prompt — this preserves prompt-cache hits on the earlier turns and lowers input costs in agentic loops. No beta header required.
• Lower prompt-cache minimum. The minimum cacheable prompt length dropped to 1,024 tokens with 4.8. Prompts that were too short to cache on 4.7 thus create cache entries without any code change.
• Refusal stop details. The stop_details object on refusals was publicly documented with 4.8. When Claude declines a request, the object describes the category of the refusal — letting applications cleanly distinguish different types of refusal and route the user accordingly.

Important for developers: the constraints from 4.7 still apply unchanged. Setting the sampling parameters temperature, top_p and top_k to a non-default value still returns a 400 error. You steer behavior via prompting, the effort parameter and adaptive thinking — not via sampling. Beyond that, according to Anthropic, 4.8 targeted three concrete areas of improvement: better handling of long agentic runs with fewer context compactions and cleaner recovery after a compaction, more reliable effort calibration across different domains, and more dependable tool triggering — meaning fewer cases where a tool call required by the task is skipped, a point some users had criticized on 4.7.

Dynamic Workflows: from assistant to agent orchestra

The innovation we consider strategically most important was hidden not in the model but in the tooling: Dynamic Workflows, at release available as a research preview in Claude Code for Enterprise, Team and Max plans. It lets you launch hundreds of parallel subagents in a single session — enough to carry out migrations across hundreds of thousands of lines of code in one pass.

This confirmed a development that had been emerging for months and that we already described as the real paradigm shift in our analysis of the Claude Code leak from March 2026: Anthropic was shifting its focus from pure model upgrades toward agentic capabilities. The guiding question for 2026 is no longer „Which model do I use?" but „How do I orchestrate multiple agents working autonomously and in parallel?". Dynamic Workflows was Anthropic's first officially shipped answer to precisely that question.

Pricing and availability

Opus 4.8 was available on all platforms from release day and remains available as a legacy model. The standard price stayed unchanged from Opus 4.7: 5 US dollars per million input tokens and 25 US dollars per million output tokens. Fast Mode, as a premium variant, cost 10 US dollars per million input tokens and 50 US dollars per million output tokens at launch; the surcharge bought up to 2.5x the output speed, not a different model.

That Anthropic kept the standard price stable despite noticeably increased performance was the real news for budget owners at the time. Since 22 September 2026 price has been the weak point of 4.8: Opus 5.5 launched at 4 and 20 US dollars, a fifth less. Anyone calculating on an Opus basis therefore calculates with Opus 5.5.

Context: Mythos, the IPO race and what came next

The release fell into a tense market phase. The race between Anthropic and OpenAI had intensified, and both houses were under the pressure of a possible IPO. In this context, Opus 4.8 was also a strategic signal: Anthropic demonstrated that it could lead in the top segment without raising the price.

Two announcements pointed beyond the model at the time. First, Anthropic planned to release lower-cost models with comparable Opus capabilities; since then at least the Opus level itself has become cheaper, with Opus 5.5 at 4 and 20 US dollars instead of 5 and 25. Second, the Claude Mythos Preview, a higher class of intelligence above Opus, was to be made available to all customers in the “coming weeks”. Mythos 5 arrived on 9 June 2026 and Mythos 5.1 on 1 September, but both only for Project Glasswing participants and aimed at defensive cybersecurity.

What this means for the mid-market

From our perspective as an agency that supports mid-market companies with AI integration, three concrete recommendations can be derived.

First: do not build new integrations on 4.8. Opus 5.5 is cheaper and, by Anthropic’s account, performs at the level of Fable 5.1 on most work. Anyone running 4.8 in production has until at least 28 May 2027 and should test the switch in a safeguarded environment against their own real tasks. Opus 5.5 does bring breaking changes, such as forced tool use returning an error; what that means for existing code is covered in our article on migrating to Fable 5.1.

Second: operationalize honesty. The roughly four-times-lower rate of unremarked code errors is only a business advantage if you build it into your own processes. Concretely: define clear escalation rules, evaluate the model's self-assessment, and automatically route uncertain results to a human. A more honest model deserves a control architecture that actually listens to its honesty.

Third: invest in agent orchestration, not in the next prompt. Dynamic Workflows shows where things are heading. The lever lies not in fine-tuning individual prompts, but in building workflows in which multiple agents work autonomously and in parallel — safeguarded, traceable, and integrated into existing systems. Whoever builds competence here now gains a lead that cannot be bought overnight.

Conclusion

Claude Opus 4.8 was not a loud generational-leap release, but a precise update with two messages: top performance was delivered at a stable price, and reliability counts more than the next benchmark record. The SWE-bench Pro score of 69.2 percent made headlines, but the roughly four times lower rate of flaws in its own code that pass unremarked is the property that turns AI from an assisting tool into a dependable part of productive processes.

For companies, the consequence remains the same one we have advocated for months: invest in robust, model-agnostic architectures and in understanding agentic workflows. Models like Opus 4.8 come and go on a quarterly cadence — the ability to embed them safely and in an orchestrated way into real business processes is the lasting competitive advantage.

Frequently asked questions about Claude Opus 4.8