How to Use Claude Opus 4.8: Coding, Honesty, and Cost

Anthropic released Claude Opus 4.8 on May 28, 2026, forty-one days after Opus 4.7 and at exactly the same rate: $5 per million input tokens and $25 per million output^[1][3]. Knowing how to use Claude Opus 4.8 well is mostly a question of what changed underneath a price tag that did not move.

Three things did. Agentic coding went up again, SWE-bench Pro climbing from 64.3% to 69.2%^[1]. Fast mode got three times cheaper. And Anthropic spent most of its announcement on something that is not a capability number at all: the model is roughly four times less likely than Opus 4.7 to let a flaw in code it wrote pass without flagging it^[1]. For anyone running agents unattended, that last one changes the risk profile more than the benchmark does.

TL;DR

• Price unchanged, capability up. $5 / $25 per million tokens, same as Opus 4.7 and 4.6^[3]. On reAPI the same model runs $4.00 / $20.00, which is 80% of Anthropic's published rate.
• The honesty gain is the headline. About four times less likely than Opus 4.7 to let its own code flaws pass unremarked, and roughly seventeen times less likely than Sonnet 4.6 to produce a dishonest summary of its own agentic work^[1].
• Fast mode dropped to a third. $10 / $50 per million tokens for roughly 2.5x output speed, against the $30 / $150 fast mode on previous Claude models^[1].
• The default effort level moved from medium to high. Harnesses that relied on the old default get deeper reasoning, higher latency, and more output tokens unless they set it explicitly^[1].
• Fixed thinking budgets were removed. Code passing budget_tokens has to move to thinking: {"type": "adaptive"}^[1].
• It does not win everywhere. GPT-5.5 still takes Terminal-Bench 2.1, 78.2% against 74.6%, and GPQA Diamond slipped slightly from Opus 4.7^[1].

What Claude Opus 4.8 is

Claude Opus 4.8 is the frontier model in Anthropic's Claude 4 family and a direct replacement for Opus 4.7 on every surface Anthropic ships to. The API identifier is claude-opus-4-8, with no date suffix^[2]. It reads up to 1M tokens in one request and returns up to 128K, or up to 300K on the Batch API with a beta header^[2].

If you want the specification-level tour, we wrote one: what is Claude Opus 4.8 covers the context window, modalities, and capability list. This guide is about the operational side, which is where the interesting decisions live.

How Opus 4.8 performs on coding and agentic benchmarks

Anthropic's headline comparison puts Opus 4.8 against Opus 4.7, GPT-5.5, and Gemini 3.1 Pro across the evaluations it considers representative of real agentic work^[1].

Opus 4.8 leads six of the seven rows. The exception is Terminal-Bench 2.1, where GPT-5.5 wins 78.2% to 74.6%^[1].

Three caveats belong next to that table, because they matter for an actual buying decision. SWE-bench Verified, reported separately, moves to 88.6% from 87.6%, a small step at a level where the benchmark is starting to saturate. GPQA Diamond went the wrong way, slipping to 93.6% from Opus 4.7's 94.2%. And these are vendor-run evaluations, so read the whole table as a directional signal rather than a neutral audit. Anthropic also reports 96.7% on USAMO 2026 and a rise to 84.3% on single-agent BrowseComp^[1].

The honest summary is "broadly ahead, not ahead everywhere."

What partners report from production

Benchmarks overstate real-world gains often enough that the early-access partner statements in Anthropic's launch note are worth more than another eval row. Several are specific enough to act on^[1].

• Databricks reported a step change in agentic reasoning at roughly 61% lower token cost than Opus 4.7. This is the one that matters for production economics, because a capability gain that also cuts spend is rare in a point release.
• Cursor reported meaningfully more efficient tool calling, completing end-to-end tasks in fewer steps at the same level of intelligence.
• Cognition, which builds Devin, reported clean tool use and instruction-following consistent enough to keep autonomous engineering workloads running unattended.
• Harvey recorded its highest score on its Legal Agent Benchmark, and the first model to clear 10% on its strictest all-pass standard.

Treat the 61% figure as a hypothesis to test on your own traffic rather than a number to put in a budget. Token cost is workload-dependent, and the default effort change described below pushes in the opposite direction.

The honesty upgrade

The most distinctive part of this release is a behavior change, not a capability number. On Anthropic's internal misaligned-behavior evaluation, scored 1 to 10 where lower is better, Opus 4.8 lands close to Claude Mythos Preview and well below both Opus 4.7 and Sonnet 4.6^[1]. In practical terms Anthropic reports it is about four times less likely than Opus 4.7 to let flaws in its own code pass unremarked, and about seventeen times less likely than Sonnet 4.6 to produce a dishonest summary of its agentic work^[1].

Two things keep that in perspective. These are rate reductions, not eliminations. A model four times less likely to hide a flaw still occasionally hides one, so human review of agentic code does not go away. And the honesty metrics come from Anthropic's own evaluations rather than third-party benchmarks, so they are the vendor's framing until independent testing catches up.

The direction still matters. In a long unattended loop, a model that proactively says it is unsure is materially safer to deploy than one that confidently ships a broken result.

Dynamic Workflows

The flagship new feature ships as a research preview in Claude Code and addresses a problem every agentic-coding user eventually hits: some jobs are too large for one agent and one context window. A framework migration across hundreds of thousands of lines. A dependency upgrade touching every service. A repo-wide refactor.

The design is orchestrator-and-workers. Opus 4.8 plans the work, splits it into independent segments, and dispatches them to hundreds of subagents running in parallel. Each subagent plans, executes, and verifies its own slice, and the orchestrator merges the verified results. Anthropic frames the repository's existing test suite as the gate for what counts as done, and describes the system as carrying a codebase-scale migration from kickoff to merge^[1].

This is an orchestration feature that the model's improved long-horizon coherence makes viable, which is also why the honesty work and the workflow work belong in the same release. Fanning out hundreds of agents is only useful if each one reliably flags what it could not finish.

Effort, fast mode, and the cost equation

Two changes here, and the first one is a silent bill increase if you miss it.

The API default effort level moved from medium to high, with xhigh and max available above it^[1]. Any harness that relied on the old default now reasons deeper, runs slower, and emits more output tokens without a line of code changing. If you are comparing Opus 4.8 cost against Opus 4.7 cost, match the effort levels first or the comparison is meaningless.

Fast mode dropped to a third of its former price. It runs the same model at roughly 2.5x output speed for $10 per million input tokens and $50 per million output, against $30 / $150 on previous Claude models^[1]. That is still double the standard rate, but it moves fast mode from "too expensive to justify" into range for latency-sensitive products.

Here is the full official rate card^[3]:

The Batch API halves both dimensions, and prompt caching makes repeated context cheap to reread at $0.50 per million tokens^[3]. Between a flat standard price, a cheaper fast tier, and an effort default that moved, the question stopped being whether Opus is affordable and became which effort level, in which mode, for which class of task.

How to use Claude Opus 4.8: migrating from Opus 4.7

Anthropic describes the move as largely non-breaking, but three changes are worth catching before production traffic flips^[1].

1. The default effort level changed from medium to high. Set it explicitly or accept deeper reasoning, higher latency, and more output tokens.
2. Fixed extended-thinking token budgets were removed. Code passing a budget_tokens value migrates to thinking: {"type": "adaptive"}.
3. The Messages API now accepts system entries mid-task without breaking the prompt cache. Not a breaking change, but a behavior to know about if your agent updates instructions mid-run.

The checklist that holds for any model upgrade: replay 20 to 50 real tasks from your production trace at matched effort levels, compare end-to-end cost and quality rather than per-token rates, re-tune prompts that pin effort or thinking budgets, and pilot Dynamic Workflows on one genuinely large job before wiring it into automation. The token-cost math is empirical. Measure it on your own workload.

Who should care right now

Teams already running Opus 4.7. The SWE-bench Pro gain plus the reported token-cost reduction make a migration pilot worth the time, provided you budget for the new high effort default and benchmark cost on real tasks first.

Teams with codebase-scale jobs. Dynamic Workflows is the reason to upgrade if you have a migration or repo-wide refactor that stalled because it was too big for one agent. Start with a single bounded job where the test suite is trustworthy.

Anyone running unattended agentic loops. The honesty improvements matter most where no human reviews each step in real time: overnight runs, autonomous pipelines, long research tasks.

Teams that should probably skip it. If you are starting fresh rather than migrating, look at Claude Opus 5 before committing to 4.8. It lists at the same $5 / $25, and on reAPI it runs cheaper than Opus 4.8 does. Opus 4.8 remains the right answer when you need a model whose behavior your evals already characterize, or when you specifically want thinking off at high effort levels, which Opus 5 no longer permits.

Calling Claude Opus 4.8 alongside your other models

Everything above is a routing problem wearing a migration costume. The effort default moved, so cost per task moved. Fast mode became viable for a class of routes it was priced out of. GPT-5.5 still wins terminal coding. Opus 5 exists now at the same list price. None of that resolves into one setting, and all of it changes again at the next release.

reAPI exposes Claude Opus 4.8 through an OpenAI-compatible /v1/chat/completions endpoint, so the client you already use for GPT and Gemini calls it too:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_REAPI_KEY",
    base_url="https://api.reapi.ai/v1",
)

resp = client.chat.completions.create(
    model="claude-opus-4-8",
    messages=[{"role": "user", "content": "Refactor this module and add tests."}],
    max_tokens=16000,
    stream=True,
)

Rates on the gateway are $4.00 per million input tokens and $20.00 per million output, which is 80% of Anthropic's published $5 / $25. The endpoint reference is at reapi.ai/docs/claude-opus-4-8, and current rates live on reapi.ai/models/claude-opus-4-8.

Swapping a model string is the easy part. Being able to swap it without a second SDK, a second key, and a second invoice is the part worth building once.

FAQ

When was Claude Opus 4.8 released?

May 28, 2026, forty-one days after Opus 4.7, at unchanged pricing^[1].

How much does Claude Opus 4.8 cost?

$5 per million input tokens and $25 per million output, with cache reads at $0.50 and the Batch API at half price^[3]. Fast mode is $10 / $50^[1].

What is the biggest thing to catch when migrating from Opus 4.7?

The API default effort level moved from medium to high. If your harness relied on the old default, you get deeper reasoning, higher latency, and more output tokens without changing any code^[1].

Does Claude Opus 4.8 beat GPT-5.5?

On five of the six rows where Anthropic compares them directly, yes. GPT-5.5 wins agentic terminal coding on Terminal-Bench 2.1, 78.2% against 74.6%^[1].

What are Dynamic Workflows?

A Claude Code research preview in which Opus 4.8 acts as an orchestrator, splits a large job into independent segments, dispatches them to hundreds of parallel subagents, and merges verified results using the repository's own test suite as the gate^[1].

Is the honesty improvement independently verified?

Not yet. The four-times and seventeen-times figures come from Anthropic's internal evaluations rather than third-party benchmarks^[1]. They are rate reductions, not eliminations, so human review of agentic output still applies.

Should I use Opus 4.8 or Opus 5?

Opus 5 lists at the same $5 / $25 and is a step-change rather than an incremental upgrade. Opus 4.8 stays relevant when your evals already characterize its behavior, or when you need thinking disabled at high effort levels, which Opus 5 rejects.

How do I call Claude Opus 4.8 with the OpenAI SDK?

Point the OpenAI client at https://api.reapi.ai/v1, set model to claude-opus-4-8, and send a normal chat request. The endpoint is OpenAI-compatible, so no SDK rewrite is needed.

Matching the effort level to the job

The useful way to read this release is that Anthropic held the price flat and spent the release on things that only show up in production: a model that flags its own mistakes, an orchestration primitive for jobs too big for one context window, and a fast mode that finally costs what latency-sensitive products can pay. The benchmark table is real but modest. The behavior change is the part that alters how much supervision an agent needs.

If you are migrating, the work is small and specific. Pin the effort level explicitly rather than inheriting the new high default, move any budget_tokens code to adaptive thinking, replay real production tasks at matched effort before trusting any cost comparison, and pilot Dynamic Workflows on one bounded job with a test suite you trust. That is how to use Claude Opus 4.8 without discovering the effort default in your invoice.

References

1. Anthropic. Introducing Claude Opus 4.8 — benchmarks, honesty evaluations, Dynamic Workflows, effort defaults, and fast-mode pricing. Retrieved July 2026 from anthropic.com/news/claude-opus-4-8
2. Anthropic. Models overview — context, output, modality, and capabilities. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/models/overview
3. Anthropic. Pricing — token, cache, and batch rates. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/pricing

Further reading

• reAPI. How to use Claude Opus 5. reapi.ai/blog/how-to-use-claude-opus-5
• reAPI. What is Claude Opus 4.8? reapi.ai/blog/what-is-claude-opus-4-8
• reAPI. Claude Opus 4.8 endpoint reference. reapi.ai/docs/claude-opus-4-8