Claude Opus 4.8 Released: Faster, Cheaper Fast Mode, Dynamic Workflows, and Honesty Improvements
Anthropic released Claude Opus 4.8 today, upgrading Opus 4.7 with benchmark improvements across coding, agentic skills, reasoning, and knowledge work. The new version is available at the same price as the previous model.
Key Features
- Effort control on claude.ai — users can now set how much effort Claude puts into a task.
- Dynamic workflows in Claude Code — lets the model tackle very large-scale problems.
- Fast mode runs at 2.5× speed and is now three times cheaper than fast mode for previous models.
Benchmark Highlights
According to the announcement, Claude Opus 4.8 is the only model to complete every case end-to-end on Anthropic's Super-Agent benchmark, beating prior Opus models and GPT-5.5 at parity on cost. On Online-Mind2Web (computer-use and browser agent tasks), it scored 84% — a meaningful jump over both Opus 4.7 and GPT-5.5. On CursorBench, it exceeds prior Opus models across every effort level.
On the Legal Agent Benchmark, Opus 4.8 is the first model to break 10% overall on the all-pass standard. Early testers also reported improvements in tool-calling efficiency (fewer steps for same intelligence), citation precision, and token efficiency on retrieval workflows.
Honesty Training
Opus 4.8 introduces explicit honesty improvements — the model is trained to avoid making unsupported claims and to flag issues with inputs/outputs proactively. This translated into higher-quality analysis and better signal-to-noise ratio in testers' evaluations.
Pricing
Opus 4.8 is available at the same price as Opus 4.7. Fast mode pricing is 3× cheaper than previous fast mode pricing. Multimodal token cost is 61% cheaper than Opus 4.7 for Genie (Databricks' AI agent).
📖 Read the full source: HN AI Agents
👀 See Also
Choosing the Best Token Provider for Your API Needs
Explore the key factors to consider when selecting a provider for tokens and APIs in AI coding and automation, based on insights from the OpenClaw community.
Go Players Disempower Themselves to AI: How Cheating Became Undetectable
The LessWrong post details how AI cheating in Go tournaments became rampant and nearly impossible to punish, using the case of Carlo Metta who used Leela 0.11 and Leela Zero to win 25 of 26 games over several seasons, with only one loss under camera surveillance.
Agent Memory Is Not a Storage Problem: It's an Authority Problem
A developer argues that agent memory fails not because retrieval misses, but because all notes return with equal authority. The fix: a graph with roles, expiration, and activation fields.
Leaked Claude Code CLI source reveals hidden features and internal flags
Analysis of leaked Claude Code CLI TypeScript source code reveals 35 build-time feature flags, including BUDDY AI pets, KAIROS persistent memory, ULTRAPLAN remote planning, and Coordinator Mode. Also found 120+ undocumented environment variables and 26 internal slash commands.