Claude Opus 4.8: Pricing, Context Window, Benchmarks, API Access

TL;DR

Last updated August 2026.

Claude Opus 4.8 shipped May 28, 2026 under the API ID claude-opus-4-8. It costs $5 per million input tokens and $25 per million output tokens, identical to Opus 4.5, 4.6, and 4.7. It has a 1M-token context window, 128K max output on the synchronous Messages API, and a January 2026 training data cutoff. Anthropic's own framing was "a modest but tangible improvement on its predecessor."

Opus 4.8 is now on Anthropic's legacy model list. Claude Opus 5 and Claude Fable 5 shipped after it, and Anthropic publishes a migration guide from Opus 4.8 to Opus 5. Opus 5 matches the $5/$25 price and the 1M context with a newer May 2026 knowledge cutoff.

What Is Claude Opus 4.8?

Claude Opus 4.8 is the Opus-tier model Anthropic released on May 28, 2026, available the same day on claude.ai, the Claude API, and Claude Code. The API ID is claude-opus-4-8, a pinned snapshot rather than an evergreen pointer. It supports adaptive thinking and does not support the older thinking.type: "enabled" extended-thinking parameter.

The release was explicitly incremental. Anthropic called it "a modest but tangible improvement on its predecessor" and said in the same announcement that "we're working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost." Simon Willison flagged that framing as unusually candid for a frontier lab release.

Three product-level features shipped alongside it: effort control in claude.ai and Cowork with high, extra, and max settings; dynamic workflows in Claude Code, which run hundreds of parallel subagents on Enterprise, Team, and Max plans; and a Messages API change that accepts system entries inside the messages array mid-task.

Opus 4.8 Pricing

Opus 4.8 costs $5 per million input tokens and $25 per million output tokens. That price has not moved across Opus 4.5, 4.6, 4.7, 4.8, and Opus 5. Fast mode costs double at $10 input and $50 output, which Anthropic describes as three times cheaper than fast mode was on previous models. Willison notes fast mode is not self-serve: access requires research preview status arranged through an account manager.

One pricing-adjacent change is easy to miss and worth real money on long agent runs. The prompt cache minimum drops from 4,096 tokens on Opus 4.7 to 1,024 tokens on Opus 4.8, so shorter prefixes become cacheable. Combined with mid-conversation system messages, you can update instructions late in a session without invalidating the cached prefix.

Context Window and Output Limits

Opus 4.8 has a 1M-token context window. Anthropic's docs put that at roughly 555,000 words or 2.5M unicode characters. Max output is 128K tokens on the synchronous Messages API. On the Message Batches API it goes to 300K output tokens with the output-300k-2026-03-24 beta header, a limit shared with Opus 5, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6.

The word count is the part that trips people up. Opus 4.7 introduced a new tokenizer that Opus 4.8 and Fable 5 both use, and the same text produces roughly 30% more tokens on it than on models before Opus 4.7. A prompt that fit a budget on Opus 4.6 bills more tokens on 4.8 for identical input.

Benchmarks Anthropic Published

Anthropic's announcement compares Opus 4.8 against Opus 4.7 and other models across coding, agentic, reasoning, and knowledge-work tasks. The named benchmarks are SWE-bench Verified, Terminal-Bench 2.1, OSWorld-Verified, CursorBench, Online-Mind2Web, Finance Agent v2, and a Legal Agent Benchmark. The scores themselves are published as an image rather than as text, so we do not reproduce individual numbers here rather than transcribe them wrong.

The result Anthropic wrote in prose is more useful than the table anyway, because it measures a failure mode benchmarks usually hide. Opus 4.8 is around four times less likely than Opus 4.7 to allow flaws in code it has written to pass unremarked. On factual accuracy, Willison reports Opus 4.8 had the lowest incorrect-rate of six models on every benchmark, and that it got there "by abstaining on questions about which it was uncertain rather than by answering more questions correctly."

That distinction matters for agent loops. A model that says "I do not know" is cheaper to supervise than one that produces a confident wrong answer three steps into a plan, even when both score the same on a pass-rate benchmark. See our SWE-bench Pro breakdown for how much scaffold design moves these numbers.

Opus 4.8 vs Opus 4.7

Price, context window, max output, tokenizer, and training cutoff are identical between the two. If you are choosing between them on a spec sheet, there is nothing to choose. The differences are behavioral and API-level.

The mid-conversation system message is the change most likely to alter how you structure a long-running agent. Opus 4.8 accepts role: "system" entries immediately after a user turn in the messages array, so you can append updated instructions late in a conversation without restating the full system prompt, and without invalidating prompt cache efficiency on the prefix you already paid for.

API Access and Model IDs

Opus 4.8 is reachable through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. The IDs differ per platform.

Two defaults to know before you benchmark cost. The effort parameter defaults to high on every Opus 4.8 surface, including the Claude API, Claude Code, and claude.ai, so you are paying for the deeper setting unless you set it explicitly. Anthropic's guidance: "On higher effort settings, Claude will think more frequently and more deeply to give better responses. On lower effort settings, Claude will respond faster and use up a user's rate limits more slowly."

Bedrock exposes global endpoints for dynamic routing and regional endpoints for guaranteed geographic data routing. Google Cloud adds multi-region endpoints between the two. Claude Platform on AWS uses first-party Claude API IDs rather than Bedrock-style IDs, and follows Anthropic's deprecation schedule instead of Bedrock's.

Opus 4.8 vs Opus 5 and Fable 5

Anthropic lists Opus 4.8 under legacy models and points new work at Claude Opus 5, with a published migration guide for moving from 4.8 to 5. Opus 5 costs the same $5 input and $25 output, keeps the 1M context and 128K max output, and carries a May 2026 knowledge cutoff against Opus 4.8's January 2026. At identical price, the newer cutoff and current support status make Opus 5 the default unless you have pinned and validated Opus 4.8 behavior you do not want to re-verify.

Claude Fable 5 sits above both at $10 input and $50 output, twice the Opus price, with a 1M context and a January 2026 cutoff. It uses the same tokenizer introduced with Opus 4.7. Anthropic describes it as its most capable widely released model and recommends Opus 5 as the starting point for complex agentic coding.

Claude Opus 4.8: Pros and Cons

Open-Weight Alternatives on Morph

Morph does not serve Anthropic models. It serves open-weight coding models on custom kernels through an OpenAI-compatible API, and the per-token gap against Opus-tier pricing is large. GLM-5.2 runs as morph-glm52-744b at $1.10 per million input and $4.10 per million output with the full 1M context, matching Opus 4.8's context window at roughly one-sixth its output price. MiniMax M3 runs as morph-minimax3-428b at $0.30 input and $1.20 output with a 256K context.

Sources: Opus 4.8 pricing from Anthropic's models overview; Morph rates from morphllm.com/pricing.

Both run against the same endpoint with the same key. Full lineup and rates are on Morph Open Source Models and pricing.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",  # Authorization: Bearer <key>
)

resp = client.chat.completions.create(
    model="morph-glm52-744b",
    messages=[
        {"role": "user", "content": "Explain this stack trace and propose a fix."},
    ],
)
print(resp.choices[0].message.content)

The tradeoff is real and worth stating flatly. Opus 4.8 and Opus 5 lead the closed frontier on Anthropic's own agentic and knowledge-work evaluations. Open-weight models close most of the coding gap at a fraction of the token cost. See GLM-5.2 and MiniMax M3 for the per-model benchmark detail, or best AI model for coding for the cross-lab comparison.

Frequently Asked Questions

What is Claude Opus 4.8?

Anthropic's Opus-tier model released May 28, 2026, API ID claude-opus-4-8. 1M-token context, 128K max output on the Messages API, January 2026 training cutoff. Anthropic called it "a modest but tangible improvement on its predecessor." It is now a legacy model, superseded by Opus 5 and Fable 5.

How much does Claude Opus 4.8 cost?

$5 per million input tokens and $25 per million output tokens, the same as Opus 4.5, 4.6, and 4.7. Fast mode is double at $10/$50, which Anthropic says is three times cheaper than fast mode on previous models. Fast mode access requires research preview status through an account manager.

What is the Opus 4.8 context window?

1M tokens, roughly 555,000 words per Anthropic's docs. Max output is 128K on the synchronous Messages API and 300K on the Message Batches API with the output-300k-2026-03-24 beta header.

What changed between Opus 4.7 and Opus 4.8?

Nothing on price, context, output limit, or training cutoff. Opus 4.8 catches flaws in its own code about four times more often, abstains rather than guessing on uncertain factual questions, accepts role: "system" messages mid-conversation, and drops the prompt cache minimum from 4,096 tokens to 1,024.

How do I call Opus 4.8 from the API?

Use claude-opus-4-8 on the Claude API, anthropic.claude-opus-4-8 on Amazon Bedrock, and claude-opus-4-8 on Google Cloud. The effort parameter defaults to high on all surfaces, so set it explicitly if you want a different level.

Should I move from Opus 4.8 to Opus 5?

Opus 5 matches the $5/$25 price, the 1M context, and the 128K max output, with a May 2026 knowledge cutoff instead of January 2026. Anthropic lists Opus 4.8 as legacy and publishes a migration guide. Stay on 4.8 only if you have pinned behavior you have already validated.

What are cheaper alternatives to Opus 4.8?

Morph serves open-weight coding models, not Claude. GLM-5.2 (morph-glm52-744b) runs at $1.10 input and $4.10 output with a 1M context. MiniMax M3 (morph-minimax3-428b) runs at $0.30 input and $1.20 output with a 256K context.