The most interesting thing about Claude Opus 4.8 is the number that didn't change. Anthropic shipped a new flagship on May 28, 2026 and left the regular price exactly where Opus 4.7 sat: $5 per million input tokens, $25 per million output. New top models almost never hold their price. This one did, and that single decision reshapes the upgrade math more than any benchmark on the launch page.
Two other facts round out what developers actually need to know. A new "fast mode" serves the same model at 2.5× the speed and, in Anthropic's words, "is now three times cheaper than it was for previous models." And Anthropic says Claude "can plan the work and then run hundreds of parallel subagents in a single session." Every number below traces to Anthropic's own announcement.
Price parity is the quiet headline
The most underrated detail is the one that didn't change. A new flagship model that costs the same as the one it replaces is unusual, and it means teams already budgeting for Opus 4.7 can move to 4.8 without re-running their cost models. Anthropic states the regular rate is $5 per million input tokens and $25 per million output tokens — identical to the prior Opus. For anything you already run on Opus, the upgrade is a model-name swap, not a pricing negotiation.
Fast mode: 2.5× the speed, a third of the old premium
"Fast mode" is the new lever. It serves the same Opus 4.8 model at "2.5× the speed," and Anthropic prices it at $10 per million input tokens and $50 per million output tokens — double the regular rate. The notable part is the comparison to history: Anthropic says fast mode for Opus 4.8 "is now three times cheaper than it was for previous models." So the speed premium still exists, but it is far smaller than the one earlier fast tiers charged. For latency-sensitive work — interactive coding, live agents, anything a human is waiting on — that changes the math on when paying for speed is worth it.
Hundreds of parallel subagents
On capability, Anthropic frames Opus 4.8 around orchestration: Claude "can plan the work and then run hundreds of parallel subagents in a single session." That is a statement about scale of delegation — one planning model fanning work out to a large pool of workers inside the same session, rather than a developer hand-managing a handful of calls. If you build agentic systems, the ceiling on "how much can one session coordinate" just moved up substantially.
The benchmark numbers Anthropic cited
Anthropic reports Opus 4.8 scoring "84% on Online-Mind2Web," a web-task benchmark, and says it is "the only model to complete every case end-to-end" on their Super-Agent benchmark. On their Legal Agent Benchmark, Anthropic says the model "has the highest score recorded" and "is the first model to break 10% overall." Read benchmarks with the usual caution — they are vendor-reported — but the throughline is agentic, multi-step tasks rather than single-shot answers.
A quieter reliability claim
One line worth flagging for anyone shipping AI-written code: Anthropic says Opus 4.8 "is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked." In plain terms, the model is more likely to catch and call out its own mistakes instead of presenting buggy code as finished. That is a meaningful property when the output goes straight into a codebase.
A new Messages API trick for long agent loops
Opus 4.8 also ships a developer-facing API change: per Anthropic, "developers can update Claude's instructions mid-task without breaking the prompt cache or routing the update through a user turn." For long-running agents, that means you can change the rules partway through — adjust permissions, budgets, or context — without paying to rebuild the cached prompt prefix or faking a user message to slip the update in. We cover the exact request shape in a companion guide.
The verdict
Adopt it. There is no reason not to. If you already run Opus 4.7, moving to 4.8 is a model-name swap with no cost re-modeling, and you inherit a more reliable code reviewer for free. Reach for fast mode only when a human is waiting on the answer, because it still doubles the per-token rate even at the lower premium. And if you build agentic loops, the two API-level changes (hundreds of parallel subagents, mid-task instruction updates that survive the prompt cache) are the parts worth your engineering time, not the headline benchmarks. Everything else you can opt into later.
Source
Written by the vybecoding.ai editorial team
Published on June 9, 2026