Claude Opus 5 lifts its test pass rate to 88.6% — and writes 2.3x more code

Code-analysis firm Sonar put Anthropic's Claude Opus 5 through its standard LLM evaluation and found the model passes more of the tests it's given — hitting 88.6% on executable coding tasks versus 82.9% for the Opus 4.8 it replaces — while generating more than twice as much code along the way.

The pass-rate gain is the headline. Across 544 HumanEval and MBPP tasks with executable tests, Opus 5 landed at 88.6%, a 5.7-point step up from Opus 4.8's 82.9%. The three Opus releases before it all clustered within a single percentage point (82.72%–83.66%), so this is a genuine jump rather than drift, and the model also finishes what it starts more reliably — missing completions fell from 1.22% to 0.83%.

On a per-line basis, the code got cleaner. Bug density dropped 14% and vulnerability density fell 20%, with Sonar reporting its most striking figure yet: blocker-level, security-impact findings tumbled 75%, from 77 to 19 per million lines of code. Exception-handling bugs fell 52% and control-flow mistakes 51%. Anthropic has framed Opus 5 as unusually good at verifying its own work while it plans; Sonar's numbers are broadly consistent with that claim.

The wrinkle is volume. For the same 4,441 tasks, Opus 5 produced 916,813 lines of code against Opus 4.8's 391,456 — a 2.3× jump — and 3.6× the output tokens (21.7 million versus 6.0 million). It split that into 2.5× as many functions, each slightly tighter, which Sonar reads as decomposition rather than sprawl. The catch is that cleaner-per-line still scales up: 2.3× more code translates to roughly 2.7× the total issues, with code-smell density up 17% and the collections-and-generics smell category more than doubling. Concurrency and threading bugs nearly doubled to become the top bug category — the same pattern Sonar saw in every GPT model it evaluated this month.

Both readings are true, and the practical takeaway is that verification load doesn't shrink — it shifts. Teams adopting Opus 5 should point automated checks where Sonar's data points: threading, cryptography configuration (its top vulnerability category, which jumped from 43 to 103 per million lines), and untyped collection handling. The volume tradeoff also complicates the model's core pitch as a cheaper alternative to Claude Fable 5 — more output means more code to review even when that output is denser with fixes. We saw the cost calculus already pulling buyers toward Opus 5 in Anthropic's Opus 5 overtakes flagship Fable 5 as buyers trade down.

What to watch: Whether Anthropic trims Opus 5's verbosity in a point release, and whether rival benchmarks start scoring code volume the way Sonar now does.

Which tradeoff matters more to your workflow — cleaner code or less of it? Tell us in the comments.

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Read more

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute