The frontier model race has been moving in fits and starts. OpenAI shipped GPT-5.5 and a new Codex line. Google pushed Gemini 3.1 Pro and a faster Gemini Flash. xAI keeps iterating on Grok. And now Anthropic has shipped Claude Opus 4.8, only 41 days after Opus 4.7, which is an unusually short release cycle for them and a clear signal about how the rest of 2026 is going to feel.
Opus 4.8 is not a flashy rebrand. The headline numbers are real (SWE-bench Pro at 69.2%, USAMO 2026 at 96.7%, GraphWalks at 1M tokens jumping from 40.3% to 68.1%), but the more interesting story is structural: effort controls you can dial per request, dynamic workflows that orchestrate hundreds of parallel subagents, a fast mode that is roughly three times cheaper than it used to be, and a measurable drop in the kind of overconfident, slightly-deceptive coding behavior that makes agents annoying to trust.
In this article, I will walk you through everything you need to know about Opus 4.8. We will cover the release details, the new effort tiers, the Dynamic Workflows feature in Claude Code, the pricing changes, the honesty and alignment improvements, two practical things you can build with it today, and an honest assessment of where it still falls short. By the end, you should have a clear mental model of when to upgrade, when to wait, and how to actually use the new controls.
What Is Claude Opus 4.8?
Claude Opus 4.8 is Anthropic's new flagship model, released on May 28, 2026 and available immediately across the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. The API model ID is claude-opus-4-8. It ships at the same headline price as Opus 4.7, keeps the 1M-token context window, and is positioned squarely at agentic coding, long-horizon reasoning, and multi-day workflows.
Here is what makes Opus 4.8 stand out:
-
Effort controls everywhere: low, high (default), extra (
xhigh), and max tiers you can pick per request, with high tuned to spend roughly the same tokens as Opus 4.7's default while doing better work. - Dynamic Workflows in Claude Code: a research preview that lets the model plan a job, spin up hundreds of parallel subagents, verify their outputs adversarially, and resume across multi-day runs.
- Cheaper fast mode: 2.5× output speed at $10/$50 per million tokens, roughly three times cheaper than the previous Opus fast mode.
- Honesty by default: a four-fold reduction in unreported code flaws versus Opus 4.7, a 0% rate of uncritically reporting flawed results (the first Claude model to hit zero on that test), and a ten-fold reduction in overconfidence.
- Big agentic coding gains: SWE-bench Pro at 69.2%, leading every published competitor on that benchmark.
- Long-context retrieval breakthrough: GraphWalks BFS at 1M tokens jumps from 40.3% to 68.1% F1, the biggest single benchmark gain of the release.
-
New API ergonomics: the Messages API now accepts
systementries inside themessagesarray, letting you update instructions mid-task without breaking the prompt cache.
Pricing and Availability
Opus 4.8 launches at the same standard rate as Opus 4.7, which keeps the upgrade math simple.
- Standard: $5 per million input tokens, $25 per million output tokens.
- Fast mode: $10 / $50 per million tokens for roughly 2.5× the output speed. Anthropic describes this as "three times cheaper than fast mode for previous models," which is the real story here: latency-sensitive workloads that were borderline on the old Opus fast mode become economically reasonable on 4.8.
- Context window: 1,000,000 tokens, unchanged from 4.7.
- Access: Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, claude.ai, Claude Code, Cowork, and (as of release day) GitHub Copilot.
Because price did not move, the upgrade decision for most teams is just a function of whether the benchmark and reliability gains help your specific workload.
Benchmark Performance
The benchmark deltas are the cleanest part of the release.
Coding
- SWE-bench Pro: 69.2% (up from 64.3% on Opus 4.7). For comparison, GPT-5.5 sits at 58.6% and the next competitor at 54.2%.
- SWE-bench Verified: 88.6% (up from 87.6%).
- SWE-bench Multilingual: 84.4% (up from 80.5%).
Opus 4.8 leads every SWE-bench variant. This is the most useful headline number for anyone building a coding agent today.
Math and Reasoning
- USAMO 2026: 96.7% (up from 69.3% on 4.7). A 27.4-point jump in one model cycle is not a normal benchmark delta. Anthropic is describing this as a qualitative change in mathematical reasoning depth, not a tuning win.
Long-Context Retrieval
- GraphWalks BFS at 1M tokens: 68.1% F1 (up from 40.3%).
- GraphWalks Parents at 1M tokens: 83.3% F1 (up from 56.6%).
This is the release's biggest relative lead. If your application leans hard on long-context retrieval (legal review, large repo navigation, research synthesis), 4.8 is a different model than 4.7 for that work.
Other Notable Scores
- HLE with tools: 57.9%
- OSWorld-Verified (computer use): 83.4%
- MCP-Atlas: 82.2%
- Finance Agent v2: 53.9%
Where It Regressed
- GPQA Diamond: 93.6% vs 94.2% on 4.7. Near-saturated benchmark, so variance at the top is expected, but worth flagging.
- Terminal-Bench 2.1: 74.6%, behind GPT-5.5's 78.2%.
- Multilingual tasks: trails Gemini 3.1 Pro and GPT-5.5 in several languages.
Take the benchmarks as directional. The honesty and reliability deltas below matter more for production use.
Effort Control: The Knob You Have Been Waiting For
Opus 4.8 exposes four effort tiers, and you can set them per request.
- Low: fast responses, minimal token use. Best for summarization, classification, and simple Q&A.
- High (default): Anthropic's "best balance" tier. Tuned to spend similar tokens to Opus 4.7's default while outperforming it on coding.
-
Extra (
xhigh): recommended for difficult tasks and long-running async workflows. - Max: maximum token depth. Reserve for quality-only priorities where you do not care about cost.
The effort knob is exposed in claude.ai and Cowork (all plans), in Claude Code via the existing effort menu, and in the API. Claude Code's rate limits have been raised to accommodate the new high default. This is the kind of control that used to live behind enterprise sales conversations, and getting it as a first-class request parameter is a quiet but meaningful win.
A practical migration heuristic: move to Opus 4.8 on the default high tier first, then sample representative tasks at xhigh to model the token-cost delta before turning it on in production.
Dynamic Workflows: Hundreds of Subagents, Multi-Day Jobs
The biggest new feature in this release is Dynamic Workflows, a research preview inside Claude Code that lets Opus 4.8 orchestrate work normally reserved for a small engineering team.
The capability, in short:
- The model plans a job.
- It spins up tens or hundreds of parallel subagents to execute pieces of it.
- Other agents adversarially try to refute the findings before they are reported.
- State is checkpointed, so jobs survive interruptions and resume across multi-day runs.
- Coordination happens outside the conversation thread, so the main session stays responsive.
The use cases Anthropic is pointing at:
- Codebase-wide bug hunts and dead-code discovery.
- Security and hardening audits.
- Large migrations: framework swaps, language ports, API deprecations.
- Verification-heavy tasks where you want independent attempts plus an adversarial reviewer.
The canonical case study is Bun's Zig-to-Rust port, run by Jarred Sumner: 750,000 lines of Rust generated, 99.8% of the test suite passing, eleven days from first commit to merge. The workflow first mapped lifetime requirements, then parallel writers generated every .rs file with two reviewers per file, and an overnight fix loop addressed a data-copy optimization. That is a real-world result, not a benchmark.
Activation
Inside Claude Code, you turn it on in any of three ways:
- Enable auto mode in Claude Code.
- Ask explicitly: "create a workflow."
- Toggle the
ultracodesetting, which sets effort toxhigh.
The first trigger shows you a preview and requires confirmation. Dynamic Workflows is available on Max, Team, and Enterprise (admin-enabled), and via the APIs.
Token Cost Warning
Workflows consume substantially more tokens than standard sessions. Plan your budgets before flipping this on in production. The same feature that finishes a quarter's worth of work in days will also spend a quarter's worth of tokens in the same window if you do not watch it.
Honesty, Reliability, and Alignment
This is the part of the release that does not show up cleanly in a benchmark table but matters most for anyone shipping agents.
Code Honesty
- Four-fold reduction in unreported code flaws versus Opus 4.7.
- Code summary honesty: fails to flag important events only 3.7% of the time.
- Uncritically reporting flawed results: 0%. First Claude model to score zero on that test.
- Overconfidence: more than ten-fold improvement over Opus 4.7.
- Lazy investigation: perfect score. Opus 4.7 gave incorrect answers 25% of the time on the same probe.
Read those as one trend: Opus 4.8 is significantly less likely to hand you a confident-sounding summary that papers over a real problem. Bridgewater Associates' testimonial in Anthropic's launch coverage explicitly calls out "Opus 4.8's tendency to proactively flag issues with the inputs and outputs of an analysis" as the differentiator.
Alignment
- Stronger prosocial behavior and user-autonomy support.
- Substantially lower deceptive and misuse-enabling behaviors than 4.7.
- Reckless or destructive actions reduced significantly.
- Overall alignment risk rated "very low."
Agentic Security Caveat
One regression worth taking seriously: prompt-injection robustness dropped. Gray Swan attack success rate climbed to roughly 9.6%, up from 6.0% on Opus 4.7. If your pipeline ingests untrusted external content (scraped web pages, user-submitted documents, third-party tool output) review your sandboxing approach before upgrading.
API Enhancements You Should Actually Notice
The Messages API now accepts system entries inside the messages array. In practice that means you can update instructions mid-task without breaking the prompt cache, which has been a real pain point for long-horizon agents.
Concrete use cases:
- Adjusting tool permissions partway through a session.
- Reallocating a token budget after a planning step.
- Injecting environment context (the user just switched repos, a long-running job finished) without restarting the whole conversation.
For long-running agentic workflows, this single change is more useful than it sounds.
How Opus 4.8 Compares to the Field
A short read on where 4.8 sits today:
vs. GPT-5.5:
- Opus 4.8 leads on SWE-bench Pro (69.2% vs 58.6%) and on GDPval-AA ELO by roughly 121 points.
- GPT-5.5 still leads Terminal-Bench 2.1 (78.2% vs 74.6%).
vs. Gemini 3.1 Pro:
- Opus 4.8 leads on SWE-bench variants and long-context retrieval.
- Gemini 3.1 Pro leads on GPQA Diamond (94.3% vs 93.6%) and on several multilingual tasks.
Read those numbers as "Opus 4.8 is the strongest agentic coding model right now, but the multilingual and terminal-harness wins are not universal."
Honest Assessment: Strengths and Limitations
Opus 4.8 is the most useful coding model Anthropic has shipped, but it is worth going in with calibrated expectations.
Where it shines:
- Agentic coding gains are real. The SWE-bench Pro lead over GPT-5.5 is large enough to matter on real workloads.
- The honesty improvements (four-fold drop in unreported flaws, zero rate on uncritical reporting, ten-fold drop in overconfidence) translate directly into less time spent double-checking the agent.
- Dynamic Workflows is a genuine shift, not a gimmick. The Bun port is the kind of result that would have been a quarter of work and is now eleven days.
- Effort control as a first-class parameter is the right primitive for cost-aware agent design.
- Fast mode at roughly one-third the previous price unlocks latency-sensitive workloads that were borderline before.
- Same pricing as 4.7 makes the upgrade decision easy.
Where to be careful:
- Prompt-injection robustness regressed (Gray Swan ~9.6% vs 6.0%). If you handle untrusted content, audit your sandbox before upgrading.
- Terminal-Bench 2.1 still trails GPT-5.5. If your workload is shell-heavy, benchmark both.
- Multilingual tasks remain a relative weakness. Gemini 3.1 Pro and GPT-5.5 win in several languages.
- Dynamic Workflows can burn tokens fast. Budget carefully or you will discover the cost on your invoice.
- Vending-Bench 2 regressed, hinting at potential issues with highly structured, multi-step transactional interactions. Worth testing if that is your domain.
- Pipelines tightly tuned to 4.7 prompts will need re-validation. The honesty improvements change how the model communicates uncertainty, which can ripple through downstream parsing.
When to Upgrade and When to Wait
Upgrade from Opus 4.7 if:
- You run agentic coding workflows and want the honesty improvements.
- Long-context retrieval is core to your product (the GraphWalks delta is the biggest gain in the release).
- You have been hitting silent failures or overconfident outputs.
- You want fast mode at the new lower price.
Stay on Opus 4.7 if:
- Your production pipeline is meticulously prompt-tuned to 4.7 and you do not have a re-validation window.
- You rely on the GPQA Diamond delta at the top of the curve.
- You ingest untrusted external content and cannot tighten sandboxing right now.
A safe default migration path: switch the model ID, leave effort at high, validate a representative sample of tasks, then experiment with xhigh for the hardest workloads.
What to Learn Next
A few directions worth exploring once you have Opus 4.8 running:
- Effort tuning per workload: profile token spend at each tier on real traffic. The right tier is almost never the same across your app.
- Dynamic Workflows playbooks: build templates for codebase audits, migrations, and dead-code sweeps. The reusable bit is the workflow shape, not the prompt.
-
Mid-task system messages: refactor your long-horizon agents to use the new in-array
systementries instead of restarting threads. - Mythos preview: Anthropic has hinted that Mythos-class models will reach general availability in the coming weeks, with the same same-price upgrade pattern. Worth tracking.
- Cheaper-with-Opus-capability tier: Anthropic also signaled cheaper models with Opus-level capability are in the works, which will reshape which tier you run by default.
Anthropic's documentation, the Claude API release notes, and the model card for Opus 4.8 are the best primary sources for any of this.
Final Thoughts
Opus 4.8 is not a rebrand and it is not a victory lap. It is a focused release that pushes on the parts of agentic coding that actually slow teams down: silent failures, overconfident summaries, single-thread workflows, and a coarse effort model. The benchmark wins are real, but the more durable shift is structural. Effort controls give you a cost dial. Dynamic Workflows give you a way to run multi-day jobs. The honesty improvements give you a model you can trust to flag its own mistakes.
The 41-day cycle from 4.7 to 4.8, the Mythos hints, and the cheaper-fast-mode pricing together signal that Anthropic is racing. That is good for everyone building on top of these models, but it also means the right play is to upgrade incrementally, measure what changed, and keep your prompts and sandboxes loose enough to absorb the next jump.
If you are building agents right now, switch the model ID, leave effort at high, test your hardest workloads at xhigh, and try one Dynamic Workflow on something you have been putting off. That is enough to see why this release matters.
