Quick Answer
Claude Opus 4.8 is Anthropic's newest flagship large language model, released to developers and enterprise customers in late May 2026. It supports a 1 million token native context window, up to 128,000 output tokens per single generation, and defaults to a high effort reasoning mode that decides for itself how much compute a prompt deserves.
The model is available now across the Claude API, Amazon Bedrock, and Google Cloud Vertex AI, with Microsoft Foundry integration rolling out. The headline upgrades are mid conversation system messages that preserve prompt cache hits for long running agents, an 84 percent score on the Online-Mind2Web browser automation benchmark, the first ever model to break 10 percent on the strict all pass legal agent benchmark, and an opt in Fast Mode that delivers up to 2.5x more output tokens per second from the same model.
Specs That Actually Matter for Builders
Anthropic has packed a lot into Opus 4.8, but a handful of numbers are the ones that will shape how developers price out a project.
- Context window. 1 million tokens supported natively by default, large enough to fit an entire mid sized codebase or a multi hundred page legal filing into a single prompt.
- Max output tokens. 128,000 tokens per single generation request, which removes the need to chain multiple calls when producing long analytical reports.
- Default execution mode. High effort. The model automatically defaults to deep, reflective processing to maximise the accuracy of its reasoning paths.
- Adaptive thinking. Opus 4.8 abandons manual user configured token budgets for extended thinking. Instead, it uses an internal self calibrating reasoning loop that decides how much compute a prompt requires.
- Prompt cache minimum. The minimum length required to trigger cost saving prompt caching has been cut to 1,024 tokens.
Mid Conversation System Messages, the Quiet Revolution
The single most interesting structural change in Opus 4.8 is the introduction of mid conversation system messages.
Previously, if an engineer needed to update an agent's instructions or inject new guardrails halfway through a long chat session, the only option was to rebuild the entire system prompt array. That process wiped out prompt cache hits, forcing the system to re-read thousands of background tokens and pushing API bills upward fast.
With Opus 4.8, developers can inject a new role: "system" message directly after any user turn in the messages array. The earlier cache stays intact, latency drops, and the cost of long running agentic loops falls significantly. For teams shipping autonomous coding agents, customer support copilots, or research assistants, this is the kind of change that quietly reshapes operational economics.
Benchmarks: Where Opus 4.8 Beats the Field
Early enterprise testing data indicates Opus 4.8 has set a new bar for autonomous operational reliability, beating prior Anthropic models and matching or surpassing rival architectures at price parity.
1. The Legal Agent Benchmark
Opus 4.8 became the first model to break the 10 percent overall barrier on the strict all pass standard for substantive legal work. In a benchmark where any single missed clause or misapplied rule fails the entire pass, breaking double digits is a real milestone. For corporate legal teams, it means contract analysis and regulatory auditing can be handed off with measurably higher statistical confidence.
2. Software Engineering and Claude Code
Integrated directly into Claude Code, Opus 4.8 powers a new research preview called Dynamic Workflows. Rather than just writing snippets, the model can map out codebase scale migrations, spin up hundreds of parallel subagents within a single session, run automated tests, catch its own syntax bugs, and execute refactors across hundreds of thousands of lines of code from initial kickoff through Git merge.
3. Browser Automation (Online-Mind2Web)
The model recorded an 84 percent accuracy rating on the Online-Mind2Web benchmark, currently the toughest public evaluation for autonomous web browser navigation. That number puts Opus 4.8 at the top of the computer use leaderboard for the moment.
Fast Mode: Same Brain, More Tokens Per Second
Deep reasoning models historically traded raw speed for accuracy. Anthropic is trying to soften that trade off with Fast Mode, a research preview that lets developers flip an API parameter to speed: "fast" and pull up to 2.5x more output tokens per second out of the same Opus 4.8 model.
Fast Mode is meant for user facing experiences where every second of perceived latency hurts engagement, like chat copilots or live code completion. For deep analytical work, leaving the model in its default high effort setting is still the right call.
Pricing and Where to Use It
Opus 4.8 is available immediately on the Claude API, Amazon Bedrock, and Google Cloud Vertex AI. Microsoft Foundry support is rolling out over the next several weeks. Anthropic has held input and output token pricing in line with the previous Opus tier while expanding the cache hit discount, which makes the new prompt caching minimum particularly valuable for any team running large multi turn agent loops.
For teams already on Claude 3.5 or earlier Opus versions, the migration path is straightforward. The API surface is backwards compatible, and the new system message rules are additive, so existing prompts continue to work without modification while you opt in to the new patterns over time.
How It Stacks Up Against the Rest of the Market
Opus 4.8 lands in a crowded field. OpenAI's GPT-5.5 still leads in some multimodal benchmarks, Google's Gemini 3.1 Pro pushes the limits on long context multimodal reasoning, and the open source community continues to close the gap on raw reasoning quality. What Opus 4.8 brings to the table is a combination of long context, agentic reliability, and price predictability that is hard to match in a single model.
For agent builders, legal tech teams, and engineering platforms, the headline question is no longer which model is theoretically smartest. It is which model can be trusted to run for hours unattended without burning through cache, drifting from instructions, or generating expensive recoveries.
The Takeaway
Claude Opus 4.8 is more than a routine point release. The combination of a 1 million token context window, mid conversation system messages that preserve prompt cache hits, an 84 percent Online-Mind2Web score, a 10 percent breakthrough on the strict legal agent benchmark, and an opt in Fast Mode for 2.5x faster output tokens makes it one of the strongest enterprise grade options on the market right now. If you are building autonomous coding agents, legal review pipelines, or long horizon research assistants, Opus 4.8 deserves to be tested on your actual workload this week.
