Claude Opus 4.8 — Anthropic's Flagship Model with Adaptive Thinking

Background and the Claude 4 Family Evolution

Anthropic, founded in 2021 by former OpenAI researchers, consistently builds its models around the Constitutional AI methodology — a design philosophy aimed at making AI systems helpful, harmless, and honest. The Claude 4 family debuted on May 22, 2025, and since then the release cadence has compressed dramatically to just a few months between iterations.

The timeline: Opus 4.1 (August 2025) focused on agentic tasks; subsequent releases — Claude Opus 4.5 (November 2025), Claude Opus 4.6 (February 2026) and Claude Opus 4.7 (April 2026) — successively strengthened coding, computer vision, and agentic behavior. Just 41 days after Opus 4.7’s launch, on May 28, 2026, Anthropic announced Claude Opus 4.8 — not as a minor patch, but as a redesigned system with the Adaptive Thinking mechanism and a one-million-token context window.

The pace reflects the brutal competition with OpenAI (GPT-5.4 and GPT-5.5) and Google (Gemini 3.1 Pro and Gemini 3.5 Flash). The AI arms race of mid-2026 is playing out in near real time.

Architecture: Adaptive Thinking and One Million Tokens

The fundamental architectural shift is the Adaptive Thinking mechanism. Earlier hybrid models required developers to manually set token budgets for reasoning. In Opus 4.8 the model itself assesses the complexity of a query and allocates reasoning automatically — the developer only controls the overall effort level.

Now the model independently assesses the complexity of each query in real time:

• For simple questions, it responds immediately, minimizing latency.
• For complex multi-step problems, it initiates an implicit reasoning loop before generating a response.
• Developers can control this via the effort parameter with values low, medium, high, xhigh, and max (default high for Opus 4.8).

Adaptive Thinking integrates automatically with tool calls — the model can "think" between sequential external API invocations, which is critical for Agentic AI workflows.

The context window stands at 1 million tokens. The maximum output tokens per turn grew from 16,000 to 128,000 — enabling complete codebases or extensive reports to be generated in a single session.

Cost Optimizations and API Innovations

The model introduces several mechanisms to reduce operational costs:

• Prompt Caching — in agentic loops, the same prefix is sent to the model on every iteration: the system prompt (instructions on how the model should behave, the list of available tools) plus the conversation history so far. Classically the model analyzes it from scratch every time. Prompt cache lets that work be preserved between calls — the first use is billed normally, subsequent ones are noticeably cheaper. Together with the new support for injecting system messages mid-conversation, the cache no longer resets when instructions change while the agent is working.
• Mid-conversation System Messages — in long-running agents (research, refactor, debug) you often want to change their behavior mid-flight: switch from fast to thorough mode, narrow the scope, warn about a budget limit. Previously every such change meant restarting the context and paying high costs. Opus 4.8 lets you push these instructions on the fly — the agent adapts on its next step, without losing what it has already done.
• Fast Mode — a variant 2.5× faster than the standard Opus 4.8. Anthropic confirms it is now three times cheaper than for previous generations (4.6, 4.7), where the speed premium was much higher. How that reduction was achieved was not disclosed.

Benchmarks and Results: Leader in Comprehensive Intelligence

On the Artificial Analysis Intelligence Index v4.0 — aggregating results from 10 rigorous evaluations (GDPval-AA, Terminal-Bench Hard, SciCode, Humanity’s Last Exam, GPQA Diamond, τ²-Bench Telecom, AA-LCR, AA-Omniscience, IFBench, CritPt) — Claude Opus 4.8 ranks first with 61 points, ahead of GPT-5.5 at xhigh (60) and high (59), and tied with Opus 4.7, Gemini 3.1 Pro and Qwen 3.7 Max at 57 points.

Coding results are strong: 69.2% on SWE-Bench Pro gives Opus 4.8 a 5-point lead over its predecessor Opus 4.7 (64.3%) and a nearly 11-point lead over GPT-5.5 (58.6%). Gemini 3.1 Pro scores 54.2% on the same evaluation. SWE-Bench Pro measures real-world problem solving on actively-maintained repositories with multi-file diffs and no public ground-truth leakage.

The GDPval-AA test, developed by Artificial Analysis on the basis of 220 tasks from OpenAI’s GDPval gold database (44 occupations across 9 economic sectors), yielded a 1890 Elo — roughly 121 Elo points ahead of GPT-5.5 at xhigh effort, implying a 66.7% pairwise win rate. Token efficiency: Databricks reports that within its Genie platform the new Opus processes PDFs, diagrams, and unstructured content at 61% lower token cost than Opus 4.7 — a direct datapoint on agentic efficiency.

On computer navigation (OSWorld-Verified) Opus 4.8 scores 83.4%, ahead of Opus 4.7 (82.3% — updated by Anthropic in the launch note), GPT-5.5 (78.7%) and Gemini 3.1 Pro (76.2%). For browser-agent work, Browserbase reports 84% on Online-Mind2Web — described as "a meaningful jump over both Opus 4.7 and GPT-5.5". On scientific reasoning, Humanity’s Last Exam with access to external tools reaches 57.9%, ahead of Opus 4.7 (54.7%), GPT-5.5 (52.2%) and Gemini 3.1 Pro (51.4%).

Practical Applications in Enterprise Environments

Software engineering gained the Dynamic Workflows feature in Claude Code. The model can architect a system plan and then spawn hundreds of parallel sub-agents within a single session, executing codebase-scale migrations using existing unit tests as the correctness criterion. Users of tools such as Devin report better situational judgment: the model asks the right questions, catches its own mistakes, and pushes back when a plan isn't sound.

Legal services: on the Legal Agent Benchmark, Opus 4.8 became the first model to break the 10% barrier under the "all-pass" standard (requiring 100% correctness across all legal steps). The CoCounsel Legal platform reports that Opus 4.8 can be safely delegated real legal work involving high responsibility toward the client — work where a law firm's error translates into real financial or procedural loss for the principal.

Data analytics and finance: on Databricks’ Genie platform, the model handles complex SQL, database querying, and visualization at a level previously out of reach. In a typical session an analyst asks in natural language ("show margin drops above 5% in Q4 by region"), and Opus 4.8 generates a SQL query against the warehouse, executes it, analyzes the result, and — crucially — notices on its own that two regions are missing December entries. Instead of silently extrapolating, it flags the gap, proposes three business hypotheses behind the drop, and suggests follow-up queries. Investment analysts highlight the high signal-to-noise ratio — earlier model generations flooded them with reports full of false positives, leaving 80% of the time on verification rather than working on the investment thesis. Opus 4.8 inverts the ratio: less output, but every flagged signal worth investigating — a direct effect of a lower hallucination rate (the lowest among the six compared models on AA-Omniscience) and better epistemic calibration.

Honesty, Hallucinations, and Sycophancy Resistance

One of the key improvements Anthropic names outright is the model’s honesty. Opus 4.8 is less likely to make unsupported claims of progress in its own work — it more often flags uncertainty about its outputs rather than confidently asserting completion on thin evidence. Anthropic’s internal alignment assessment finds that Opus 4.8 has substantially lower rates of misaligned behavior (deception, cooperation with misuse) than Opus 4.7 — matching Anthropic’s best-aligned model to date.

Hallucinations: on the AA-Omniscience benchmark, Opus 4.8 has the lowest incorrect-answer rate among the six compared models — achieving this mainly by abstaining when it doesn’t know a fact, rather than guessing. In code the effect is even more tangible: Opus 4.8 is four times less likely to let through a hidden bug than Opus 4.7. These are situations where the model generates code that "looks correct", compiles, but contains a subtle logical defect (off-by-one, race condition, wrong conditional branch). Earlier generations silently accepted such bugs inside agentic loops, leading to compounding technical debt. Opus 4.8 stops and flags doubt more often.

It’s worth noting, though, that Gemini 3.1 Pro simply knows more. On factual recall — how many facts a model can correctly retrieve from memory — Gemini still leads Opus 4.8. These are two different quality dimensions: Opus hallucinates less often, but when asked about a hard fact it more often honestly admits it doesn’t know — rather than providing an answer Gemini would simply recall from memory. For use cases that reward broad erudition (scientific research, fact-checking, encyclopedic queries) Gemini still wins.

Pricing, Availability, and Competitive Pressure

Opus 4.8 launched with no price increase over 4.7: $5.00 per million input tokens and $25.00 per million output tokens. The API model ID is claude-opus-4-8.

Fast Mode costs $10/$50 per million tokens — 2× the standard rate instead of 6× in previous versions. The model has been available across all major cloud platforms since launch day.

What this means for the market

Opus 4.8 closes a cycle of four releases in under twelve months (4.5 November 2025 → 4.6 February → 4.7 April → 4.8 May 2026). Anthropic kept pricing flat versus 4.7, added a 1-million-token context window, raised the per-turn output cap to 128,000 tokens, and made Fast Mode cheaper — meaning more capability for the same rate card. That shifts the cost calculation for teams that previously used Opus only for expensive, critical tasks. Agentic workflows, repo-wide refactors, and long research sessions become economically viable as everyday tooling.

Opus 4.8’s real edge does not sit in any single benchmark. It comes from three properties that are hard to measure individually: a roughly fourfold drop in unflagged code defects versus 4.7, the lowest hallucination rate among the six compared models on AA-Omniscience, and improved honesty calibration confirmed by Anthropic’s own alignment assessment. For use cases where a model error has real cost — legal, financial, production code — that combination matters more than a one-point edge on a leaderboard.

At the same time Opus does not win everything. Gemini 3.1 Pro remains stronger on raw factual recall, GPT-5.5 stays close at xhigh effort, and on specific tasks (Finance Agent v2 — Gemini 3.5 Flash) the competition still leads. As of late May 2026: Anthropic has the broadest model for agentic work, but the high-end LLM market remains a multi-player race, and the gap between the leaders has narrowed to a handful of Elo points.

Sources:

• Anthropic: anthropic.com/news/claude-opus-4-8 — Opus 4.8 launch announcement
• Anthropic: anthropic.com/claude-opus-4-8-system-card — Claude Opus 4.8 System Card (PDF, full benchmark data)
• Artificial Analysis: artificialanalysis.ai — Intelligence Index v4.0 (public leaderboard)
• Anthropic: claude.com/blog/introducing-dynamic-workflows-in-claude-code — Dynamic Workflows in Claude Code