Claude Opus 5: what changes vs Opus 4.8 and when to use it (2026)

In 60 seconds: Anthropic launched Claude Opus 5 on July 24, 2026. It keeps Opus 4.8 pricing at $5/MTok input and $25/MTok output, adds a 1M-token context window, enables thinking by default, and scores 97.00% on SWE-bench Verified. It fits agentic coding and standard knowledge work where Fable 5’s 95% score does not justify twice the price. Look elsewhere if you need self-hosting; choose Fable 5 if you need Anthropic’s highest capability regardless of cost.

Claude Opus 5 arrived two months after Opus 4.8, which launched on May 28, 2026. The quick release cycle makes the practical questions more useful than the announcement: what changed, what stayed the same, and which workloads justify a migration?

Disclosure: I started testing Opus 5 when Claude API access opened on July 24. It was only one day old when I wrote this, so this is not a long-term review. The observations below include prompts I ran against Opus 5 and Opus 4.8 in parallel, alongside the published data.

What changes vs Opus 4.8

Anthropic describes Opus 5 as close to Claude Fable 5 at half the price. The release has three changes that matter in practice.

1. Stronger agentic coding

Opus 5 leads SWE-bench Verified at 97.00%, ahead of GPT-5.6 Sol (96.20%) and Fable 5 (95.00%). Compared with Opus 4.8 at 88.60%, that is an increase of about 8.4 points. On SWE-bench Pro, which uses longer and messier problems, Opus 5 reaches 79.2%; Fable 5 reaches 80.0% and Mythos 5 leads at 80.3%. The difference is most relevant in multi-hour sessions and navigation across large repos.

2. Thinking enabled by default

Through Opus 4.8, extended thinking was opt-in: you had to pass thinking={"type": "enabled", "budget_tokens": N} on each call. In Opus 5 it is enabled by default, with an adaptive budget based on prompt complexity. Multi-step tasks therefore get extra reasoning without an API change.

3. More stable long-running agents

Anthropic reports improvements on multi-hour agentic sessions with little oversight. Combined with the 1M context window, this lets a session keep more of a large codebase in context without compaction.

Anthropic also reinforced its cybersecurity safeguards after conversations with government regulators about dual-use concerns. In practice, users coming from Opus 4.8 may see more rate limiting and more refusals on ambiguous prompts.

Specs and pricing

Spec	Value
Model ID	claude-opus-5
Context window	1M tokens (default and max)
Max output	128k tokens
Thinking	Enabled by default (adaptive budget)
Input pricing	$5 / million tokens
Output pricing	$25 / million tokens
Availability	Claude API, Amazon Bedrock, Claude Pro/Max/Team/Enterprise

Pricing is identical to Opus 4.8. At these rates, Opus 5 costs half as much as Fable 5 ($10/$50 MTok) and one quarter as much as Mythos 5, the tier above Fable.

Availability by plan:

• Claude Max: Opus 5 becomes the default model. If you had Opus 4.8 configured, migrate to claude-opus-5 to get default thinking and better coding.
• Claude Pro: Opus 5 is the strongest model available in the Pro tier.
• Claude Team and Enterprise: available with the governance you already have configured.
• Claude API and Amazon Bedrock: model claude-opus-5 since July 24.

How to use Opus 5 with the Anthropic SDK

The Python SDK is unchanged; only the model ID is different. This is the minimum request:

import os
from anthropic import Anthropic

client = Anthropic(api_key=os.environ.get("ANTHROPIC_API_KEY"))

message = client.messages.create(
    model="claude-opus-5",  # used to be "claude-opus-4-8"
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "Refactor this module to use async/await instead of callbacks."}
    ],
)

print(message.content[0].text)

If you were using opt-in extended thinking (with thinking={"type": "enabled", ...}), you can drop the block: Opus 5 already thinks by default. If you want explicit budget control, the parameter still works:

response = client.messages.create(
    model="claude-opus-5",
    max_tokens=16000,
    thinking={
        "type": "enabled",
        "budget_tokens": 10000,  # explicit override of the adaptive default
    },
    messages=[{"role": "user", "content": "Solve this scheduling problem..."}],
)

Streaming works the same as before, and the thinking block shows up in response.content with type="thinking" before the final text block. If you had logic reading thinking blocks, nothing changes.

For more SDK and endpoint details, the official Anthropic docs are up to date.

What the benchmarks show

Public data as of July 25, 2026, comparing the strongest models available:

Coding (SWE-bench Verified):

Model	Score
Claude Opus 5	97.00%
GPT-5.6 Sol	96.20%
Claude Fable 5	95.00%
Kimi K3	93.40%
GPT-5.6 Luna	93.00%
Claude Opus 4.8	88.60%
Grok 4.5	86.60%

Opus 5 leads this coding benchmark, but the gap to Fable 5 (1.8 points) and GPT-5.6 Sol (0.8 points) is small. In agentic coding, differences of that size rarely produce a clear productivity gain across ordinary tasks.

SWE-bench Pro (longer and messier): Mythos 5 leads at 80.3%, Fable 5 at 80.0%, Opus 5 at 79.2%. Opus 5 is 0.8 points behind Fable 5, and the gap starts showing on multi-hour tasks where problems accumulate.

Frontier-bench v0.1 (general reasoning): Opus 5 reaches a 43.3% pass rate. I don’t have exact Fable 5 and Mythos 5 numbers on this benchmark at the close of this post, so this comparison stops with the published Opus 5 result.

What Opus 5 still doesn’t reach vs Fable 5:

• Pure reasoning on very long problems (Fable 5 maintains coherence better on 4+ hour sessions)
• Professional knowledge work where the gap between “97% correct” and “99% correct” is material (legal, medical research, fine financial analysis)
• Inference speed under massive load (Fable 5 has more optimized infra)

On agentic coding and standard knowledge work, Opus 5 is indistinguishable from Fable 5 for 95% of workloads. The gap shows up only in the last 5%, where Fable 5’s extra cost starts to make sense.

When to pick Opus 5, and when not to

Pick Opus 5 if:

• Your workload is agentic coding, refactoring, or standard knowledge work
• You want to keep Opus 4.8’s cost ($5/$25) but step up a tier of capability
• You’re on Claude Max and want the updated default model
• You need 1M context without compaction to pass entire codebases
• Anthropic’s ecosystem (SDK, integrations, governance) works for you

Pick Claude Fable 5 if:

• Your workload is pure multi-step reasoning on very long sessions (4+ hours)
• Professional knowledge work where the last 1-2% of accuracy is material
• Cost isn’t a concern and you want Anthropic’s absolute ceiling
• You need aggressive prompt caching (Fable 5 has cache hits at $1/MTok)

Pick MiniMax M3 if:

• You need self-hosting for compliance, cost, or data control
• Your product is inherently multimodal (video, audio + text)
• You want to avoid vendor lock-in to Anthropic
• Token volume justifies operating your own infra (dedicated H100)

Pick Kimi K3 if:

• You want open weights with more long-horizon coding maturity than M3
• Your workload is primarily text (you don’t need M3’s native multimodality)
• You can wait until July 27 for open weights without paying API

Pick GPT-5.6 Sol if:

• Your stack is already on OpenAI and switching creates friction
• You need OpenAI-ecosystem-specific features (Assistants API, GPT Store, etc.)

What Opus 5 changes for LATAM teams

For developers and companies in LATAM, Opus 5 matters for three concrete reasons:

1. Better price/performance ratio on managed API. LATAM still pays USD per token like the rest of the world, but the gap between $5/MTok (Opus 5) and $10/MTok (Fable 5) shows up in bills at medium volume. For teams of 5-20 engineers running coding agents, migrating from Opus 4.8 to Opus 5 gets them ~8 extra SWE-bench points at no cost change.

2. Claude Max becomes the default model. If many Spanish-speaking developers are on Claude Max (likely, given the relatively accessible pricing versus going direct on Fable 5), Opus 5 reaches them automatically. The only action is to verify that the updated model ID is in use.

3. Default thinking reduces friction. In LATAM, where many teams adopt AI later than in the US, opt-in thinking was a barrier: you had to know it existed, read the docs, configure it. With Opus 5 it’s on by default, which lowers the learning curve for teams just starting with agents.

What Opus 5 doesn’t solve for LATAM:

• Data compliance. Opus 5 is still a managed API on Anthropic’s servers. If you need data in-region (local regulations), self-hosting with M3 or Kimi K3 is still the only path.
• Cost at high volume. Self-hosting with open weights is still 30-50% cheaper than any managed API past ~50M tokens/month.

Before migrating, run the same representative prompts on Opus 4.8 and Opus 5, then compare completion quality, token use, and the number of corrections your team makes. The unchanged price makes that test easier to justify than a broad migration based on benchmark scores alone.