Claude Opus 4.8

Note: Claude Opus 5 was released on July 24, 2026, replacing Opus 4.8 as the Opus-tier flagship. Opus 5 is the same price ($5/$25) and beats 4.8 on every published benchmark. The main reason to stay on 4.8 is its marginally better hallucination calibration.

Claude Opus 4.8 was, as of June 2026, the overall intelligence leader — the highest score any released model had posted on the Artificial Analysis (AA) Intelligence Index (61.4), narrowly ahead of GPT-5.5 (60.2). It tops the hardest agentic coding (SWE-bench Pro 69.2%), computer use (OSWorld-Verified 83.4%) and professional knowledge work (GDPval-AA 1,890 Elo), and pairs that with the most careful hallucination calibration of any frontier model.

Released May 28, 2026 as a point upgrade over Opus 4.7, the headline isn’t raw scores — those moved a few points — but judgment. Anthropic reports Opus 4.8 is the first Claude to score 0% on uncritically reporting flawed results, and roughly four times less likely than 4.7 to let a code defect slip through unflagged. For teams, that’s the real story: Opus 4.8 is the model you reach for when a wrong answer is expensive — production code, legal and financial analysis, multi-hour agent runs.

The catch is cost. At $5 / $25 per 1M tokens it’s the priciest model in Anthropic’s own lineup and many times the price of open-weight coders that score within a few points on the hardest benchmarks. A new, cheaper Fast Mode and a Dynamic Workflows preview in Claude Code change the cost-and-latency math, but the base rate card is unchanged.

Benchmarks

Opus 4.8’s gains over 4.7 are concentrated in agentic and knowledge-work tasks rather than raw Q&A. The numbers below are Anthropic-reported as of June 2026; treat them as a dated snapshot.

A few caveats worth knowing:

• SWE-bench Pro vs Verified. Pro filters out tasks where models show signs of memorization, so 69.2% there is harder-won than the 88.6% on Verified. The ~11-point lead over GPT-5.5 on Pro is Opus 4.8’s clearest coding advantage.
• Terminal-Bench is where GPT-5.5 wins. On the common Terminus-2 harness GPT-5.5 scores 78.2% to Opus 4.8’s 74.6%; on its own Codex CLI harness GPT-5.5 reports 83.4%. If terminal-native agents are your workload, see GPT-5.5.
• Knowledge work. Anthropic puts Opus 4.8 ~121 Elo ahead of GPT-5.5 on GDPval-AA, its measure of real-world professional tasks.

Pricing and cost

The base rate card has held since Opus 4.5. What moved in 4.8 is Fast Mode and prompt caching.

Practical implications:

• Caching is the biggest lever. At ~10% of input, cache reads make large, stable system prompts and tool definitions cheap to reuse across a session. The lower 1,024-token minimum means even small prompts are worth caching.
• Fast Mode is now economical. Because Fast Mode dropped to one-third of its previous price, latency-sensitive flows (interactive agents, IDE-style loops) can run the same frontier model rather than dropping down to Sonnet.
• Routing still matters. For high-volume or low-stakes work, Gemini 3.1 Pro (~$2 input) or open weights like DeepSeek V4-Pro cost a fraction per token. Reserve Opus 4.8 for the steps where reliability pays for itself.

Capabilities and what changed in 4.8

• Hallucination calibration. The flagship change. Opus 4.8 is markedly less likely to assert unverified results or pass over defects silently — the trait that makes it the default for high-stakes agent runs.
• Effort control. Opus 4.8 uses adaptive thinking with an effort parameter that defaults to high on every surface (API and Claude Code included). Lower it to trade depth for speed and cost. On Claude.ai, users get a matching effort slider.
• Dynamic Workflows (research preview). In Claude Code, Opus 4.8 can write an orchestration script and run parallel subagents — up to 16 concurrent and 1,000 total per run — to tackle codebase-scale tasks in a single session.
• Computer use. Strong desktop control (OSWorld-Verified 83.4%, ~84% on Online-Mind2Web) for browser- and GUI-driven automation.
• Long context. A reliable 1M-token window with high needle-in-haystack recall, plus up to 128K tokens of output (300K via a Batch API beta header).

Safety and alignment

Anthropic ships frontier models under its Responsible Scaling Policy (RSP), which gates capability behind tiered AI Safety Levels (ASL). Opus 4.8 is deployed under ASL-3 safeguards — the same tier as the rest of the Opus 4 line — with classifier-based monitoring for the highest-risk misuse categories. The next tier up, the Mythos-class Fable 5, went generally available on June 9, 2026 behind a new classifier layer — and notably, its cyber/bio/distillation fallbacks route to Opus 4.8.

What matters in day-to-day use:

• Calibration is the safety story. The headline 4.8 change is honesty under uncertainty: Anthropic reports 0% on uncritically reporting flawed results and roughly 4× fewer silently-missed defects than 4.7. In practice the model surfaces “I’m not sure” instead of confabulating — the trait that makes it safer to put in a loop.
• Trained with Constitutional AI + RLHF. Refusals are tuned to be less over-cautious than earlier Claudes on benign requests while still declining genuinely harmful ones.
• Agentic hardening. Because Opus 4.8 acts on web content, files and tools, prompt-injection and jailbreak resistance were a focus. No frontier model is immune, so keep human approval gates on irreversible or high-impact actions (payments, deletes, deploys).
• Data handling. API inputs and outputs are not used to train Anthropic models by default, and zero-data-retention is available to eligible accounts — relevant for legal, health and finance workloads.

You can see how these priorities show up in the model’s own instructions in the Claude Opus 4.8 system prompt section below.

Use cases and worked examples

Opus 4.8 earns its premium on tasks where a wrong answer is expensive or a run is long enough that small error rates compound. Three patterns where it’s the right call:

1. Long-running coding agents

This is the flagship use case. With Dynamic Workflows (research preview) in Claude Code, Opus 4.8 can write its own orchestration script and fan out parallel subagents — up to 16 concurrent and 1,000 total per run — to take on codebase-scale refactors, migrations and test-fixing in one session. The high SWE-bench Pro score (69.2%) plus its defect-flagging calibration means it’s less likely to “fix” a test by quietly breaking behaviour.

import anthropic

client = anthropic.Anthropic()

resp = client.messages.create(
    model="claude-opus-4-8",     # AWS Bedrock: anthropic.claude-opus-4-8
    max_tokens=8000,
    effort="high",               # default; lower to "medium"/"low" to cut cost + latency
    messages=[
        {
            "role": "user",
            "content": "Refactor this module to remove the global state, and flag anything you are not certain about.",
        }
    ],
)
print(resp.content[0].text)

2. High-stakes analysis (legal, finance, research)

When the cost of a confident-but-wrong answer is high — contract review, financial modelling, due diligence, scientific literature synthesis — Opus 4.8’s calibration is the differentiator. Prompt it to mark assumptions and low-confidence claims explicitly; it complies well, turning a black-box answer into an auditable one.

3. Computer-use automation

With OSWorld-Verified at 83.4% (~84% on Online-Mind2Web), Opus 4.8 is strong at driving real desktop and browser UIs — form filling, data extraction across legacy apps, multi-step workflows that have no API. Pair it with approval gates so the agent pauses before anything destructive.

For high-volume or low-stakes variants of these jobs, route the bulk to Sonnet 4.6 or a cheaper model and reserve Opus 4.8 for the hardest steps (see prompting tips below).

Prompting and best practices

• Set effort deliberately. It defaults to high on every surface, which is the most capable and the most expensive/slow setting. Drop to medium or low for routine work — you keep the model’s judgment at a fraction of the latency.
• Cache aggressively. At ~10% of input, prompt-cache reads make stable system prompts and tool definitions nearly free to reuse across a session; the 1,024-token minimum means even small, repeated context is worth caching.
• Escalate, don’t default. Start traffic on Sonnet 4.6 and escalate only the hardest steps to Opus 4.8. Paying $5/$25 for every turn of an agent loop is the most common way to overspend.
• Lean on its calibration. Explicitly ask it to “flag assumptions and anything you’re unsure about.” Opus 4.8 follows this reliably, which is exactly what you want in high-stakes flows.
• Give clean tool schemas. For agents, clear, well-described tools beat clever prompting. Use Dynamic Workflows for codebase-scale parallelism, and keep approval gates on irreversible actions.
• Structure long context. Recall across the 1M window is strong, but putting key instructions near the start and end and chunking source material still improves reliability.

Limitations and known gotchas

• Terminal/CLI agents go to GPT-5.5. On the common Terminus-2 harness GPT-5.5 leads 78.2% to 74.6%. If your workload is terminal-native, GPT-5.5 is the better default.
• Text and image in, text out only. Opus 4.8 does not accept audio or video and does not generate images, audio or video. For omnimodal I/O see GPT-5.5 or Gemini 3.1 Pro.
• It’s the price ceiling. At $5/$25 it’s the most expensive model in Anthropic’s lineup and 5–12× open-weight coders that score within a few points on the hardest benchmarks. Output-heavy jobs at $25–$50 per 1M tokens add up fast.
• Latency at default effort. Deep thinking adds latency; use Fast Mode or a lower effort for interactive, latency-sensitive loops.
• Calibration can read as over-caution. The same honesty that prevents confabulation can mean more hedging and clarifying questions — tune it back with prompting where you want decisiveness.
• Knowledge cutoff is January 2026. Supply current facts, docs or tools for anything more recent.
• Context caps vary by surface. Microsoft Foundry tops out at 200K tokens vs the full 1M elsewhere.
• Dynamic Workflows is a preview. Expect rough edges; don’t put it on a production-critical path without guardrails.

How to access Claude Opus 4.8

Opus 4.8 is available everywhere on day one. Knowledge cutoff is January 2026.

On Microsoft Foundry the context window is capped at 200K tokens; everywhere else it’s the full 1M. Cost controls available across surfaces include prompt caching, the Batch API, context trimming and model routing to Sonnet/Haiku.

Migrating from Opus 4.7

For most callers the upgrade is a model-string swap — the API shape and the $5/$25 base rate card are unchanged. The gains are in judgment, not interface. Still, a few behaviours moved:

Tokenizer note: the tokenizer that raised effective costs by up to ~35% landed at Opus 4.7, and 4.8 keeps it — so token counts are stable on a 4.7 → 4.8 move. Only re-estimate spend if you’re coming from Opus 4.6 or earlier.

How Claude Opus 4.8 compares

The June 2026 frontier is a tight cluster — there is no single best model, only a best model for your task. A fuller breakdown lives in our 2026 LLM landscape guide; the short version:

• vs GPT-5.5: Opus 4.8 wins overall intelligence and SWE-bench Pro; GPT-5.5 wins terminal agents and accepts audio/video. Output is cheaper on Opus ($25 vs $30).
• vs Gemini 3.1 Pro: Gemini is the value play (~$2 input) and natively multimodal; Opus pulls ahead on the hardest coding and on calibration.
• vs open weights (DeepSeek V4-Pro, Grok 4.3): these land within a few points on coding at 5–12× lower price. Opus 4.8’s premium buys reliability and the top of the leaderboard, not a runaway capability gap.

The Claude 4.8 family

Opus is the top tier; most production traffic should start on Sonnet and escalate to Opus only for the hardest steps.

Above Opus sits Anthropic’s Mythos-class tier: Fable 5 went generally available on June 9, 2026 (after an 18-day export-control suspension, it returned globally July 1), with its cyber/bio/distillation queries falling back to Opus 4.8. The tier first surfaced in the leaked Opus system prompt as the “Claude Mythos Preview” under an internal “Project Glasswing” (see below). Below Opus, Sonnet 5 replaced Sonnet 4.6 on June 30 and closes much of the agentic gap at a fraction of the price.

Version history

The Opus 4.x line has shared the same $5 / $25 rate card since Opus 4.5. Opus 4.7 introduced a new tokenizer that raised effective costs by up to ~35% for some migrations. Opus 4.8 (May 28, 2026) keeps that rate card but adds the headline calibration/honesty gains, a Fast Mode that’s three times cheaper than before, the Dynamic Workflows preview, and the top spot on the AA Intelligence Index — the first time a Claude model has led that index outright.

Best for

• Production coding and long-running software agents
• Legal, finance and analysis where a wrong answer is expensive
• Computer-use automation (OSWorld 83.4%)

Pros & cons

Frequently asked questions

Sources

• Anthropic — Introducing Claude Opus 4.8
• Anthropic — Claude models overview
• Anthropic — Claude Opus
• Anthropic — Pricing
• Anthropic — Responsible Scaling Policy
• Anthropic — Claude Code