Opus 4.8 Released: Is it Good? (Real Tests and Review)

Anthropic released Opus 4.8, roughly 40% faster than 4.7, and the creator calls it a genuine comeback after 4.7 was pretty much trash.

Honest marketing, finally

Anthropic downplayed 4.8 instead of hyping it. The model card leans on the word honesty repeatedly, and the release notes flag a four-times improvement in misalignment behavior, with user benchmarks showing it's almost five times more honest. The creator interprets this as Anthropic learning its lesson after 4.7: they want to underell and over deliver now.

Still behind on agentic coding

Opus 4.8 scores 74.6% on EnnC coding's Sbench Pro, compared to 78% for GPT 5.5. The creator confirms this is consistent with what they saw during 4.7, when Opus fade from view and developers migrated to GPT 5.5 on Codex. The 4.8 model card itself admits it still not so good on a thing, which the creator credits as unusually transparent.

Real-world tests: games and 3D

Tested on a futuristic space shooter with physics based movement, particle effects, combo system, and wave based enemy spawning as a single prompt, Opus 4.8 shipped a working build with secondary fire, a super bolt on hold, and heat management. Output was definitely better than the same prompt on Kimmy, and most certainly better than the DeepSeek V4 Pro version, though the physics effect is still lacking behind — no screen shake, unlike the Qwen 3.7 Max test.

A 3D Chinese building test showed Opus still drops the roof (the roof is on the floor), a bug it shares with Kimmy. GPT 5.5 gets the roof right but loses the columns. On a Minecraft-clone (Voxal World) and the Pelican riding a bike test, GPT 5.5 has more flare — the creator says Opus is now more boring, the inverse of its old reputation.

Why it feels smoother

A key underrated change: 4.8 calibrates response length to task complexity instead of using a fixed verbosity. The creator says this is very underrated and is the main reason 4.8 feels better than glory days of 4.6 despite lower benchmark scores.

Caveats and cost

Don't run heavy agent teams immediately. The creator triggered suspicious activity on Opus by burning 80% to 90% of a 5-hour token limit in 30 minutes and got banned. Anthropic reset usage limits so users can test, but the SpaceX $1B server lease deal (now tuned to 6 months) is what will determine whether 4.8 is a real turning point or another watered-down release. Bottom line: I do say it's back, but the creator is reserving judgment until Mythos ships.