Claude Opus 4.8 — What Is New in Anthropic Flagship

Anthropic released Claude Opus 4.8 today, May 28, 2026 — a flagship update that changes nothing about how you call the model and quite a lot about what comes back. Pricing stays at $5 per million input tokens and $25 per million output, the 1M-token context window stays at standard rates, and the API surface is identical to Opus 4.7. The upgrade is, quite literally, a model-ID swap.

TL;DR:

• Claude Opus 4.8 (claude-opus-4-8) launched May 28, 2026 at the same $5/$25 per MTok as Opus 4.7, with 1M context at standard pricing.

• No API breaking changes: same adaptive-thinking-only surface, no sampling parameters. Migrating is a one-line model-ID change.

• The gains are in clearer, warmer writing, long-horizon agentic work, and code review/debugging — Anthropic reports 83.4% on Terminal-Bench 2.1 and 82.3% on OSWorld-Verified.

• New: mid-session system messages for steering agents, and fast mode support at $10/$50 per MTok — three times cheaper than fast mode was on 4.7.

• Update: since June 9, Claude Fable 5 sits above Opus in the lineup. Opus 4.8 remains the default recommendation for demanding work at half Fable’s price.

Update, July 12, 2026: this article was published on release day and has been updated since — most notably with the section on where Opus 4.8 sits now that Claude Fable 5 exists above it.

Claude Opus 4.8 at a glance

Spec	Claude Opus 4.8	Claude Opus 4.7
API model ID	claude-opus-4-8	claude-opus-4-7
Release date	May 28, 2026	Early 2026
API price (in/out per MTok)	$5 / $25	$5 / $25
Context window	1M tokens, standard pricing	1M tokens
Max output	128K tokens	128K tokens
Thinking	Adaptive only	Adaptive only
Fast mode	Yes — $10 / $50 per MTok	Yes — $30 / $150 per MTok
Mid-session system messages	Yes	No
Breaking API changes	None — drop-in model-ID swap

What is new in Claude Opus 4.8 vs Opus 4.7

Clearer, warmer writing

The most immediately noticeable change is prose quality. Opus 4.7 had a tendency toward stiff, over-qualified answers on open-ended writing tasks; 4.8 reads more like a competent colleague and less like a policy document. In my first sessions the difference showed up in emails, documentation, and explanation-heavy answers — fewer hedges, better paragraph rhythm, and a willingness to commit to a recommendation. If you rejected Opus for writing work before, it is worth a re-test.

Stronger long-horizon agentic work

Anthropic’s release announcement reports gains across agentic benchmarks, including 83.4% on Terminal-Bench 2.1 and 82.3% on OSWorld-Verified. The numbers matter less than what they represent: 4.8 stays coherent deeper into long sessions. In multi-hour Claude Code runs I have watched it recover from a failed approach and re-plan rather than looping — the failure mode that most often forced a human back into the loop on 4.7. The release also landed alongside a dynamic-workflows capability in Claude Code for very large-scale tasks.

Better code review and debugging

Code review is where 4.8 earns its keep for me. It is measurably better at spotting the bug that matters in a diff — the off-by-one in error handling, the cache key that ignores a parameter — rather than producing a list of ten stylistic nitpicks. Debugging follows the same pattern: it forms a hypothesis, asks for the right log line, and revises when the evidence contradicts it, instead of pattern-matching to the most common cause.

Mid-session system messages

Opus 4.8 supports system messages injected mid-conversation, not just at the start. For agent builders this is a genuinely useful primitive: you can tighten constraints, hand off between agent roles, or update instructions partway through a long session without restarting it and losing context. Previously this required awkward workarounds like stuffing instructions into user turns, which models weight differently.

Fast mode support — at a third of the old price

Opus 4.8 supports fast mode, the research-preview option that serves the same model at up to 2.5× output tokens per second. The launch news is the price: fast Opus 4.8 costs $10/$50 per MTok, down from $30/$150 on 4.7 — three times cheaper. That moves fast mode from “demo budget” to something you can defensibly run for latency-sensitive agents. We cover setup, rate limits, and the fallback pattern in our full Claude fast mode guide.

Effort control arrives on claude.ai

Alongside the model, Anthropic shipped a user-facing control on claude.ai for how much effort Claude puts into a task — the consumer cousin of the API’s effort parameter. In day-to-day use it is the difference between a quick sanity-check answer and a genuinely worked-through one, without switching models. It defaults sensibly, but for anything you would previously have prefaced with “think carefully about this,” turning effort up on Opus 4.8 is now the honest way to get that behavior instead of hoping a prompt incantation does it.

No API breaking changes: migration is a model-ID swap

If you are already on Opus 4.7, migration is one line:

- "model": "claude-opus-4-7",
+ "model": "claude-opus-4-8",

Everything else carries over. Opus 4.7 already removed the sampling parameters (temperature, top_p, top_k) and moved to adaptive thinking without budget_tokens, so 4.8 introduces no new rejections — the official what’s-new page confirms the API surface is unchanged. If you are jumping from something older than 4.7, you inherit those two migrations; the parameter cleanup is the same one described in our other model coverage, and it takes minutes, not days.

One practical note from my own migration: because outputs are qualitatively different (warmer prose, different code-review emphasis), re-run your evaluation suite even though nothing “breaks.” Prompts tuned to compensate for 4.7’s stiffness can overshoot on 4.8.

Pricing: unchanged, and the 1M context stays standard

Opus 4.8 costs exactly what 4.7 costs: $5 per million input tokens, $25 per million output. Just as importantly, the 1M-token context window is billed at standard pricing — there is no long-context surcharge for requests that push past 200K tokens, which is where competing platforms often add multipliers. Prompt caching, batch processing, and the other standard discounts all apply as before.

Unchanged pricing plus measurable capability gains is the quiet story of this release: your cost per successful task goes down even though your cost per token does not. On claude.ai, Opus 4.8 is available in the model picker on paid plans, and it works in Claude Code within your plan’s usage budget — alongside a new claude.ai control over how much effort Claude puts into a task.

Update: where Opus 4.8 sits now that Fable 5 exists

Added June–July 2026. Twelve days after this release, on June 9, Anthropic launched Claude Fable 5 — the first model of the Claude 5 family and the company’s most intelligent generally available model, positioned explicitly above Opus. So is Opus 4.8 yesterday’s flagship? In practice, no. Fable 5 costs $10/$50 per MTok, double Opus 4.8 on input and output, and its June suspension-and-restoration saga made single-model dependencies look risky. Opus 4.8 has settled into the role Anthropic seems to have designed it for: the default recommendation for demanding work, with Fable 5 reserved for the tasks where the extra intelligence is worth 2× the price. Late June also brought Claude Sonnet 5, which presses Opus from below on bounded coding tasks. For the full ladder — Haiku to Fable — see our Claude models comparison, and for how the Opus/Sonnet decision plays out in practice, the Opus vs Sonnet deep dive.

The pricing continuity also matters for anyone doing capacity planning. Opus 4.7 remains an active API model at the same $5/$25, so there is no forced-march deadline: you can run 4.7 and 4.8 side by side during evaluation, split traffic, and cut over when your numbers say so. (In hindsight — a note added with the July update — that is a noticeably gentler upgrade path than the Sonnet 5 transition a month later, where a tokenizer change and an intro-pricing deadline put a clock on the decision.) Here, the only cost of waiting is leaving the better model unused.

Should you upgrade to Claude Opus 4.8?

• From Opus 4.7: yes, immediately. Same price, same API, better model. There is no scenario where staying on 4.7 wins — and note that fast mode on 4.7 is being wound down in favor of 4.8’s much cheaper fast tier.
• From Sonnet-tier models: upgrade the tasks, not the account. Route long-horizon agent runs, hard debugging, and high-stakes writing to Opus 4.8; leave bounded, high-volume work on Sonnet where the economics favor it.
• To Fable 5 instead: only with a reason. If your hardest tasks still fail on Opus 4.8, Fable 5 is the escalation path. Start with Opus; escalate on evidence, not on launch-day excitement.

What people get wrong about Opus 4.8

• “No breaking changes means no testing needed.” The API contract is identical; the model behavior is not. Re-run your evals — especially prompts that were tuned around 4.7’s quirks.
• “Fast mode makes it smarter.” Fast mode serves the same weights at higher output speed. Intelligence is identical; only tokens per second and price change.
• “Fable 5 made Opus obsolete.” Fable sits above Opus at double the price. For most demanding work, Opus 4.8 remains the sensible default; Fable is the escalation tier.
• “The 1M context costs extra.” It does not — long-context requests on Opus 4.8 bill at the standard $5/$25.
• “Benchmark deltas will match my workload.” Anthropic’s reported results (like 83.4% on Terminal-Bench 2.1 and 82.3% on OSWorld-Verified) are directional. Measure on your own tasks before re-routing traffic.

FAQ

Bottom line

Claude Opus 4.8 is the easiest flagship upgrade Anthropic has shipped: same price, same API, better everywhere it counts, with mid-session system messages and affordable fast mode as genuine new tools. Swap the model ID, re-run your evals, and route your hardest work here — and if you want the long-term hands-on verdict beyond release coverage, our evergreen Claude Opus review tracks how it holds up in daily use. For everything else in the lineup, start at the guide section.