Introducing Claude Opus 4.8

TL;DR

Claude Opus 4.8 is Anthropic’s GA closed-source frontier model, an incremental upgrade over Opus 4.7 at unchanged pricing ( $5/$ 25 per M input/output tokens, fast mode now 3× cheaper at $10/$ 50). Two product surfaces matter for research watchers: (a) a new effort control on claude.ai exposing the inference-compute knob (low → high → extra/xhigh → max) to all end users, and (b) dynamic workflows in Claude Code that lets Opus 4.8 plan, dispatch hundreds of parallel subagents in a single session, and verify outputs before reporting back — explicitly framed as codebase-scale migration over hundreds of thousands of lines of code with the test suite as the bar. Anthropic also previews a new model class, Claude Mythos Preview, under Project Glasswing cyber-safeguards, positioned as higher-intelligence than Opus.

Key claims

Pricing is held flat at $5/M input and$ 25/M output tokens for regular usage; fast mode is $10/$ 50 and is now 3× cheaper than for previous models [§Availability, §Introducing Claude Opus 4.8].
A user-facing effort control ships on claude.ai across all plans, sitting next to the model selector, with four tiers (low / high / extra ≡ xhigh in Claude Code / max) — Opus 4.8 defaults to “high” as the best quality/UX balance [§Also launching today, §A note on effort].
At default “high” effort on coding tasks, Opus 4.8 spends a similar number of tokens as Opus 4.7’s default but with better performance [§A note on effort].
Dynamic workflows in Claude Code (research preview, Enterprise/Team/Max plans) lets Claude plan a task, run hundreds of parallel subagents in a single session, verify their outputs, and report back — explicitly designed for codebase-scale migrations over hundreds of thousands of lines of code, gated by the existing test suite [§Also launching today].
The Messages API now accepts system entries inside the messages array, so harnesses can update permissions, token budgets, or environment context mid-task without breaking the prompt cache or routing the update through a synthetic user turn [§Also launching today].
Opus 4.8 is reported to be ~4× less likely than Opus 4.7 to allow flaws in its own code to pass unremarked, framed as an honesty/self-flagging improvement rather than raw capability [§Collaborating with Opus 4.8].
The alignment assessment summarizes Opus 4.8 as reaching “new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest,” with rates of misaligned behavior (deception, cooperation with misuse) substantially lower than Opus 4.7 and similar to the internal best-aligned Claude Mythos Preview [§Collaborating with Opus 4.8].
Partner-supplied benchmark / capability claims (no shared external axis, transcribed from the testimonial block):
- Hebbia: only model to complete every case end-to-end on their Super-Agent benchmark, beating prior Opus models and GPT-5.5 at parity on cost [§Collaborating with Opus 4.8].
- Cursor: exceeds prior Opus models across every effort level on CursorBench; tool calling uses fewer steps for the same intelligence [§Collaborating with Opus 4.8].
- Harvey: highest recorded score on their Legal Agent Benchmark; first model to break 10% overall on the all-pass standard [§Collaborating with Opus 4.8].
- Browserbase / computer-use partner: 84% on Online-Mind2Web, a “meaningful jump” over both Opus 4.7 and GPT-5.5; positioned as the strongest computer-use / browser-agent model tested [§Collaborating with Opus 4.8].
- Databricks Genie: multimodal reasoning over PDFs/diagrams/unstructured content reported at 61% cheaper token cost than Opus 4.7 [§Collaborating with Opus 4.8].
- Devin / Cognition: fixes the comment-verbosity and tool-calling issues observed with Opus 4.7 [§Collaborating with Opus 4.8].
Footnote evals: Terminal-Bench 2.1 scores reported using the public Terminus-2 harness; GPT-5.5’s reported Terminal-Bench 2.1 score with the Codex CLI harness is 83.4% [§Footnotes]. OSWorld-Verified evaluation methodology was changed and Opus 4.7’s score was updated to 82.3% as part of the rerun [§Footnotes]. Finance Agent v2: Gemini 3.5 Flash scores 57.9% [§Footnotes].
A higher-capability model class Claude Mythos Preview is in limited use under Project Glasswing for cybersecurity, awaiting stronger cyber safeguards before general release; Anthropic says it expects to ship Mythos-class models to all customers “in the coming weeks” [§What’s next?].
API model id: claude-opus-4-8; available on the Claude API and partner clouds [§Availability].

Method

This is a closed-source product release — no architecture, training data, or post-training recipe is disclosed. What is described:

A continuation of the Opus 4.x line; the named differentiators are (a) judgment/self-flagging on agentic tasks, (b) the exposed effort-tier control end-to-end, and (c) parallel-subagent dispatch via dynamic workflows.
The honesty/uncertainty-flagging improvement is described as the model being more willing to surface uncertainty about its own work and less likely to make unsupported claims — quantified only at the level of the “4× less likely to allow code flaws to pass unremarked” claim and the alignment-team summary.
Two infrastructure-facing surfaces appear for the first time on the wiki:
1. Dynamic workflows: a Claude Code harness mode where the model plans, fans out to “hundreds of parallel subagents” within a single session, then verifies their outputs against the test suite before merging — the same shape as PARL-style learned parallel orchestration but at a product layer, with no claim that the orchestrator policy is RL-trained.
2. In-thread system messages: mid-conversation system entries inside the messages array, designed to update permissions / budgets / context without breaking prompt-cache reuse or requiring a synthetic user turn. This is a small but concrete primitive for long-running agent harnesses that previously had to either invalidate the cache or smuggle context via user turns.
Cyber safety: continues the Opus 4.7 pattern of differential capability reduction + runtime misuse detection + a verification program (Project Glasswing), now extended to gate the next-class Mythos model from general release.

Results

Benchmark surface in the post is the same shape as the Opus 4.7 release — a partner-testimonial block plus a small footnote table. No shared external axis, so direct numerical comparison to the open MoE cluster (Qwen3-Coder-Next Technical Report, Step 3.5 Flash) remains impossible. Notable numbers:

Online-Mind2Web 84% as the strongest computer-use / browser-agent number Anthropic has shipped, exceeding both Opus 4.7 and GPT-5.5 on the same axis per the partner statement. Closes part of the gap on the OSWorld / computer-use axis that the Agentic Software Engineering page flagged as missing from open filings.
~4× reduction in unremarked code flaws vs Opus 4.7 — Anthropic’s first quantitative claim about self-flagging at this magnitude on an Opus release.
Databricks Genie 61% cheaper token cost than Opus 4.7 on multimodal document reasoning — first cost-side claim co-reported with a capability claim on an Anthropic release page, mirroring the inference-cost-on-the-headline-axis trend that Step 3.5 Flash surfaced on the open side.
Fast mode 3× cheaper than prior models at $10/$ 50 per M input/output tokens, with 2.5× decoding speed — a serving-side cost reduction at the same quality, not a new capability tier.

Safety: alignment-team summary attributes “new highs” on prosocial traits and substantially lower misalignment rates than Opus 4.7, with the comparison anchor being internal Claude Mythos Preview (positioned as best-aligned to date). Full numerics deferred to the Claude Opus 4.8 System Card.

Why it’s interesting

For the Agentic Software Engineering thread, dynamic workflows is the first time a closed-frontier release ships a product-layer parallel-subagent orchestrator — the same shape as Kimi K2.5’s PARL (Kimi K2.5: Visual Agentic Intelligence) and Cursor’s swarm result (Cursor: agent swarm discovers novel solution to Problem Six of the First Proof challenge), but framed at the harness layer rather than as a trained policy. The 84% Online-Mind2Web number is the first Anthropic-side datapoint that meaningfully advances the computer-use axis that Introducing GPT-5.3-Codex put on the closed-frontier benchmark quadruple. For Inference-Time Scaling, the effort knob is now end-user-facing across all plans — the closed-source frontier has fully productized the “spend more tokens to think more” axis that Introducing Claude Opus 4.7 only exposed via the developer-facing xhigh tier. The Mythos-class teaser keeps the Opus → Mythos capability ladder visible without committing a release date, mirroring the Glasswing-gated cyber-safeguard pattern that 4.7 established. The 4× self-flagging improvement is the most concrete honesty/calibration claim Anthropic has put on an Opus release page so far; whether it shows up on the broader honesty/hallucination benchmarks tracked by Why Language Models Hallucinate is the natural follow-up read.