Is Claude Opus 5 Bad? Pricing Impact & Test Results

Updated July 28: A developer’s article titled “Opus 5 is a really bad model” has drawn more than 320,000 views on X and reached Hacker News. It reports serious failures in Claude Code—but the evidence is one user’s experience after several hours, not a controlled model evaluation.

The useful conclusion is narrower: Opus 5 may behave differently enough in tool-heavy coding sessions that migration needs production-shaped testing. The live table and chart above use our current pricing dataset, while the calculator lets you model your own token volume.

What the Opus 5 criticism says

The author reports that Opus 5 ignored CLAUDE.md and rule files, avoided reading or listing files, guessed at workspace contents, preferred shell commands over Claude Code’s native tools, and sometimes edited without adequate verification.

They also describe missed document sections, context loss across repeated /compact runs, and reluctance to use sub-agents. The critique attributes some behavior to Claude Code’s shorter system prompt and specific instructions about error reporting, blocked work, and agent delegation.

Those are consequential claims for autonomous coding. A model that saves tokens but causes a bad edit, a repeated prompt, or a lost instruction can cost more per accepted result. However, the post does not provide a reproducible task set, comparison logs, sampling settings, or failure rates. Its claim that Anthropic’s benchmarks are fraudulent is therefore not established by the evidence shown.

One Hacker News commenter reported the opposite experience: no notable regression from Opus 4.8, better agent-to-agent communication, and stable long sessions with extensive project instructions. That disagreement reinforces the need for workload-specific tests.

What our current Opus 5 test found

AI Pricing Guru added Opus 5 to Labs on launch day. In the latest published 49-task run on July 28, it completed every request with zero API errors and scored 47/49, or 95.92%. Opus 4.8 scored 49/49.

The two Opus 5 misses were exact-format instruction checks. It returned invalid for a valid JSON string, then wrapped the correct pricing metric in bold Markdown instead of returning only the requested plain text. Those bounded failures do not establish a broad regression, but they mean the launch-day 49/49 result is no longer the current score.

Our larger warning is efficiency. Opus 5 produced 2,051 completion tokens versus 621 for Opus 4.8 in the same narrow suite. Its measured run cost was $0.077435 versus $0.041685, making cost per correct answer about 1.94 times higher despite identical list rates.

The two results are not contradictory. Our suite checks bounded answers; the viral report concerns long-horizon Claude Code behavior, tool selection, injected instructions, and compaction. Neither test alone answers how Opus 5 performs on your repository.

Pricing impact

Anthropic launched Opus 5 on July 24 as a same-rate replacement for Opus 4.8: $5 input, $0.50 cache-hit input, and $25 output per 1M tokens. The official model ID remains claude-opus-5, with a 1M-token context window and up to 128K output. Anthropic also says it approaches Fable 5 on key coding evaluations at lower task cost. Those claims make Opus 5 attractive—but only if production completion rates hold.

For coding agents, list price is the floor. Total cost also includes output length, retries, tool calls, failed edits, CI runs, human review, and time spent restoring lost context. The viral critique matters because each reported failure increases one of those hidden costs.

Sonnet 5 remains the economical default for routine traffic. Fable 5 may justify its premium when its higher completion rate avoids expensive failed loops. Compare the full lineup on our Anthropic pricing page and alternatives on the OpenAI pricing page.

What developers should do now

Run an Opus 5 canary before changing the default. Use real repositories and score instruction retention, required-file reads, tool choice, verification steps, accepted patches, retries, compaction fidelity, output tokens, and human correction time.

Keep the prompt and Claude Code version fixed when comparing Opus 5 with Opus 4.8 or Sonnet 5. Test long sessions separately from short tasks, because compaction and context handling are central to the new complaint.

Add explicit checks for important instructions and require validation before edits land. Keep a fallback model available until Opus 5 has passed representative work. Use the AI token calculator for the rate-card estimate, then choose on cost per accepted task.

Bottom line

The viral “bad model” verdict is a valuable bug report, not a universal benchmark. It identifies plausible Claude Code integration failures that can overturn token savings, while our bounded test shows Opus 5 can still be accurate and reliable on a different task class.

Existing Opus users should test the upgrade, not assume it. Teams already succeeding with Sonnet 5 should keep it as the default and escalate only the work where Opus 5 proves its premium through fewer failures or better outcomes.

Sources: the developer’s Opus 5 critique on X, the Hacker News discussion, Anthropic’s Opus 5 announcement, Anthropic’s official pricing documentation, and the live AI Pricing Guru API dataset.