Skip to main content
Back

Anthropic's Claude Opus 4.8 Cuts Coding Bugs 4x in 41 Days

Claude Opus 4.8 lands in Cursor with better agentic coding scores, 4x fewer silent bugs, and a 3x cheaper fast mode at the same base price.

  • Claude Opus 4.8 is now available in Cursor, with better agentic coding scores and 4x fewer silent code bugs than Opus 4.7.
  • SWE-bench Pro score jumps from 64.3% to 69.2%, putting it 10+ points ahead of GPT-5.5 (58.6%) on real-world coding tasks.
  • Fast mode is now 3x cheaper ($10/$50 per million tokens vs. $30/$150), with Databricks reporting 61% lower token costs in production.
  • No long-context pricing multiplier: the full 1M token context window bills at the same flat rate as short prompts.
  • Dynamic workflows in Claude Code (Max/Team/Enterprise) lets Opus 4.8 spawn hundreds of parallel subagents for codebase-scale migrations and audits.
  • Alignment improvements match Claude Mythos Preview levels; Anthropic says Mythos-class models for general release are coming "in the coming weeks."

Anthropic's Claude Opus 4.8 is now available inside Cursor, arriving just 41 days after Opus 4.7. The update keeps the same price tag but ships with measurably better agentic coding performance, a dramatically cheaper fast mode, and a model that is far less likely to silently let bugs slide past you. Anthropic calls it "a modest but tangible improvement" -- but the numbers and early partner reports tell a more interesting story.

The Opus 4.7 problems it quietly fixes

Opus 4.7 had a rough reception. The Devin team specifically called out that Opus 4.7 had comment-verbosity and tool-calling issues -- and these were not minor complaints: the verbosity issue inflated developer bills, and the tool-calling issue caused agentic coding pipelines to fail mid-execution. Opus 4.8 directly addresses both.

  • 4x fewer silent bugs: Anthropic reports the new model is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked."
  • Fewer tool-calling steps: On CursorBench, Opus 4.8 uses fewer steps for the same intelligence, meaning the token-per-task cost drops without sacrificing pass rates.
  • Less verbose output: The model uses 35% fewer output tokens per task, directly fixing the verbosity problem, and completes tasks in 15% fewer turns.

The practical upshot: the pricing stayed flat at $5/$25 for standard mode, which means the effective cost per task actually decreased because the model uses fewer tokens to accomplish the same work.

Benchmark numbers worth knowing

On SWE-bench Pro -- a coding benchmark using real-world repositories with multi-file diffs -- Opus 4.8 lands at 69.2%, almost 5 points clear of Opus 4.7 (64.3%) and over 10 points ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%). The harder the benchmark variant, the wider the gap gets.

The math reasoning jump is even more striking. Opus 4.8 scored 96.7% on the USAMO 2026 math benchmark, up from 69.3% on Opus 4.7 -- a 27.4 percentage point gain in a single 41-day release cycle, the biggest single-cycle math improvement in Opus history. USAMO problems require multi-step proof construction, so this signals a structural improvement in how the model reasons through hard, multi-step problems -- the same kind of reasoning that matters in complex agentic workflows.

For browser-agent and computer-use tasks: Opus 4.8 scores 84% on Online-Mind2Web, which is a meaningful jump over both Opus 4.7 and GPT-5.5. And on the aggregate Artificial Analysis Quality Index, Opus 4.8 sits at 61.4, ahead of GPT-5.5 at 60.2 -- the first time a Claude model has held the top position on this index since GPT-5.5's release.

What Cursor specifically gets

Cursor describes Opus 4.8 as "a meaningful jump over Opus 4.7 on CursorBench," excelling at autonomous multi-step work: it holds intent across long sessions, self-corrects when it hits friction, and writes production-ready code without hand-holding. They recommend the high thinking variant for best results.

The Cursor docs highlight four specific strengths:

  • Autonomous and self-directed: Opus 4.8 drives multi-step tasks to completion without losing track of the goal, even across large codebases and long conversations.
  • Creative reasoning: It approaches problems from unexpected angles, explores alternative solutions, and produces more inventive code than its predecessor.
  • Strong planning: It maps out work before executing, catches edge cases early, and builds coherent architectures across many files.
  • Reliable tool use: It calls tools purposefully, chains tool results into follow-up actions, and adapts when tool output surprises it.

Pricing inside Cursor

Cursor plans include two usage pools. Opus 4.8 draws from the API pool, which charges at per-token rates. Individual plans include at least $20 of API usage each month (more on higher tiers). The base rate is $5/M input and $25/M output tokens -- unchanged from Opus 4.7.

The bigger change is in fast mode. A fast mode tier (claude-opus-4-8-fast) is available for roughly 2.5x faster output. It requires Max Mode and bills at $10/M input and $50/M output tokens -- 3x lower than Opus 4.7 fast mode. Databricks ran Opus 4.8 Fast Mode in their Genie AI agent and reported 61% cheaper token costs versus Opus 4.7.

One notable pricing detail: all Opus 4.8 prompts bill at the base per-token rates, including when you use Max Mode and context goes above 300k. There is no separate long-context multiplier for Opus 4.8 -- up to 1M tokens at the same rates.

Dynamic workflows: the bigger picture

Opus 4.8 ships alongside a new Claude Code feature called dynamic workflows -- and the two are designed to work together. Available in research preview, it allows Claude to take on even bigger tasks in Claude Code. Claude can plan the work and then run hundreds of parallel subagents in a single session, and with Opus 4.8 the agents can run for even longer.

Practical use cases for this combination include:

  • Codebase-wide migrations across hundreds of thousands of lines of code
  • Security audits across entire services
  • Bug hunts and language ports at repository scale
  • Deep research and multi-document analysis pipelines

Dynamic workflows are available on Max, Team, and Enterprise plans.

What partners are saying

The partner quotes from Anthropic's launch page paint a consistent picture: better judgment, less noise, more reliability. Claude Opus 4.8 has noticeably better judgment -- in Claude Code, it asks the right questions, catches its own mistakes, pushes back when a plan isn't sound, and builds up confidence around complex, multi-service explorations before making big changes.

On Hebbia's Super-Agent benchmark, Claude Opus 4.8 is the only model to complete every case end-to-end, beating prior Opus models and GPT-5.5 at parity on cost -- delivering powerful reliability for agent products in translation, deep research, slide-building, and analysis.

On the legal side, Opus 4.8 delivers the highest score recorded on Harvey's Legal Agent Benchmark, and is the first model to break 10% overall on the all-pass standard. That benchmark evaluates the full range of tasks junior associates perform: reading dense contracts, identifying risks, cross-referencing regulatory frameworks, and drafting legal memoranda.

Alignment improvements that matter for agents

One underreported aspect of this release is the alignment story. Anthropic's Alignment team concluded that Opus 4.8 "reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user's best interest," and that misaligned behavior rates are substantially lower than Opus 4.7 -- similar to their best-aligned model, Claude Mythos Preview.

That's the first time a generally-available Claude has been benchmarked at Mythos-class alignment levels. For long-running autonomous agents that operate unattended, a model that is less likely to deceive or cooperate with misuse is not just a nice-to-have -- it is a prerequisite for production deployment.

What's next on the roadmap

Anthropic is signaling that Opus 4.8 is a stepping stone, not a ceiling. Anthropic confirmed that Mythos-class models -- with "even higher intelligence than Opus" -- will be available to all customers "in the coming weeks." Currently, Claude Mythos Preview is limited to cybersecurity work under Project Glasswing. The company is developing safety guardrails to enable broader release.

For Cursor users today, the upgrade path is simple: if you are on Opus 4.7 via API or Claude Pro, the upgrade is free -- same model string, same price. The model is available now via claude-opus-4-8 in the Claude API and selectable directly inside Cursor's model picker.