AI Weekly Malaysia

livenerf is an append-only, pre-registered benchmark built to test whether a frontier model quietly degrades after launch, and it started the clock on Claude Opus 5.5, released 2026-09-22. Day 1 ran 2026-09-24 22:10 UTC, roughly 2.5 days after launch, and it now samples once a day for 30 days: days 1-10 form the baseline, then two 10-day windows, so the first possible drift call lands around 2026-10-24 and the first Results row after day 20. It runs through headless Claude Code (claude -p) on a Claude Max subscription with no API key, using frozen prompts, a pinned CLI version, exact graders and raw logs, built on the UK AI Security Institute's Inspect framework with error bars per Anthropic's 'Adding Error Bars to Evals'; the repo has 366 stars and the Hacker News thread has 343 points and 147 comments.

Why: If you ship anything on Claude models, this is the closest thing to a day-0 baseline anyone has published, and it says plainly that sampling parameters are gone and thinking can't be turned off, so reproducibility has to come from pinning the CLI version, freezing prompts and keeping raw logs. The concrete decision: pin your model and CLI version in a file the way this repo does, log raw outputs now, and treat any post-launch quality claim as unproven until there are thousands of samples with error bars - not vibes. Note the timeline: no drift verdict exists before roughly 2026-10-24, so anything claiming Opus 5.5 was 'nerfed' before then is speculation.

Hillel Wayne's Buttondown post What TLA+ can and can't check responds to Boris Cherny's claim that Opus used TLA+ to find race conditions in code, pushing back on the idea that formal methods will solve agentic software development. It walks through what TLA+ can express, including behaviors as state sequences, the temporal operators [] always, P' next, and <> eventually, plus invariants and action properties, while promising to focus on properties TLA+ cannot even express. The excerpt ends mid-explanation of action properties and stutter-invariance, and the Hacker News thread had 222 points and 47 comments.

Why: If you use coding agents like Claude Code or Opus for bug-fixing, do not treat a TLA+ run as a turnkey correctness guarantee: the author notes you still need a property to verify, and correct designs do not automatically translate into correct code. Teams should decide who writes and reviews the invariants or properties before trusting agent-generated fixes.

AMD is buying World Labs for $8.2B — a price the roundup says is known only because AMD is public — less than two years after World Labs' 2024 founding, on the back of its spatial-intelligence models and its SceniX acquisition for robotics simulation. World Labs says Atlas, trained from scratch, predicts the next camera view from 2D images and outperforms specialized models on the long-standing computer-vision problem of sparse reconstruction by combining generative models with multiview geometry, with interest cited in robotics RL environments, scene generation, and real-estate/design/construction reconstruction. The same roundup reports Anthropic shipped Claude Sonnet 5.5 a week after Opus 5.5, claiming 30%+ faster and up to 30% cheaper than Sonnet 5 for most work, with early independent evals placing it at or near Opus 5.5 and Anthropic positioning it for 'well-scoped everyday tasks like fixing bugs and quickly iterating on features.'

Why: The Sonnet 5.5 claim is the one you can act on now: if you default to Opus for bug fixes and feature iteration, a 30% cost cut at near-Opus eval scores is worth re-measuring on your own repo before your next billing cycle. The World Labs deal is the opposite — a large acquisition and a capable-sounding model, but the text gives no Atlas API, pricing, license, or availability, so there is nothing to build on yet; treat it as a signal that 3D/scene reconstruction is consolidating into big-chip money, not as a tool you can adopt this week.

Anthropic released Claude Sonnet 5.5, which per Anthropic "runs 30%+ faster, and costs up to 30% less for most work" while priced the same as Sonnet 5, and in Simon Willison's hands-on tests it beat Sonnet 5 on every benchmark and came close to Opus 5.5 on some coding tasks. Sonnet 5.5 is now the model behind the free tier on claude.ai, which Willison notes makes Anthropic's free offering more capable than ChatGPT's free tier running Luna 5.6. He also reproduced an Opus 5.5 failure mode: at "max" thinking effort the model burned 128,000 tokens (~$1.28) and failed to produce an SVG, while "xhigh" effort produced output in 41 seconds for 5.74 cents; Haiku 5.5 is still promised "in the coming weeks".

Why: If you pay for Sonnet-tier API calls, the same price now buys a model that is roughly 30% faster and cheaper to run, and Willison reports it nearly matching Opus 5.5 on coding tasks — a concrete reason to re-run your evals before defaulting to a pricier model. If you prototype on free tiers, claude.ai's free tier now serves Sonnet 5.5 rather than a weaker small model, so the WebGL-pelican-style prompt he tested is a free way to gauge output quality before spending. Set a thinking-token ceiling: his "max" run spent $1.28 and 128,000 tokens and still returned nothing.

Anthropic's docs page for prompting Claude Opus 5.5 describes behavioral differences from Opus 5 and gives harness patterns for them. The one hard number in the text: Opus 5.5 generates output tokens more than 30 percent faster than Opus 5 and tends to finish the same task with fewer tokens, and existing Opus 5 prompts are said to work unchanged. The page is organized as a symptom index (effort calibration, thinking-disabled prompts, unattended agents that stall after reporting progress, stop_reason "refusal", silent long agentic turns, multi-app context, multiagent time signals, pasted text being followed as instructions, complex visual inputs, generic frontend output) and points to a separate migration guide for four breaking API changes from Opus 5. The excerpt is cut off before the actual capability details and before those four breaking changes are listed.

Why: If you already ship on Claude Opus 5, the two things that force action are the four breaking API changes and the documented failure modes: agents that stop partway after a progress update, silent long agentic turns, and stop_reason "refusal" responses all have named fixes here rather than guesswork. The 30 percent faster output tokens and fewer tokens per task is the only cost/latency claim in the text, so treat it as a reason to re-measure your own token spend after swapping the model ID, not as a reason to swap blindly. Because the excerpt is truncated, you cannot see the four breaking changes or the effort-calibration guidance from this text alone - open the migration guide before changing anything.

Anthropic introduced mods for Claude Code: small TypeScript functions that hook into events Claude Code emits (tool calls, permission prompts, UI draws) to run before, after, instead of, or wrapping them. A mod can rewrite prompts before they reach the model, block/retry tool calls, approve or deny permissions, redact secrets from tool output, and add or replace UI elements in the CLI, desktop app, or both. Mods ship inside plugins and are explicitly not sandboxed — they run with the same machine access as Claude Code itself, so the post says to install only from trusted sources.

Why: If your team runs Claude Code, this is a new install surface: mods arrive as plugins and execute unsandboxed with the same access as Claude Code, so the practical decision is whether to allow third-party mods at all and who reviews them. The useful capability to note is prompt/tool-call interception — you can now redact secrets from tool output before Claude reads it, or block and retry specific tool calls, without waiting on Anthropic to ship a feature. Hooks previously could not rewrite events, draw UI, or replace features; mods can, which changes what you'd build in-house versus install.

A 1h32m Latent Space episode with Anthropic's Thariq Shihipar framed as a catch-up on Claude Code's recent releases: Opus 5.5, a Plugins portal, and Cloud Sessions/Claude Projects arrived 'last week,' Sonnet 5.5 shipped the day of recording, and Claude Mods — user-extensible behaviour for Claude Code, with a linked cheatsheet and GitHub issue — is the item the show flags for special attention. The intro also lists Anthropic's claimed numbers (largest-ever raise in May at $47B ARR, $65B ARR in July, IPO target of $2T with ~$100B ARR estimated for end-2026) and earlier Claude Tag, Fable 5 and Mythos 5.1 launches, plus a Boris Cherny post about a community-built Tetris-in-Claude mod (603K views, 312 replies). The excerpt itself contains no benchmarks, pricing, or migration detail — it is a launch list plus a podcast pointer.

Why: If you run Claude Code in a team, the two things that change your setup are Claude Mods and the Plugins portal: they are a distribution channel for shared agent behaviour, so the decision is whether to package your existing prompts/config as a mod/plugin or keep it as a private repo. Separately, Opus 5.5 and Sonnet 5.5 landing roughly a week apart means any model version pinned in your CI or agent config will likely need bumping soon, so pin deliberately and note what you'd have to re-test. Treat the ARR, IPO and model-launch figures as vendor framing — the text gives no independent verification.

Anthropic sales development leader Carl Johnson describes a 'buying agent' built on Claude Managed Agents (beta) that takes prospects from the Contact Sales form through to checkout, reportedly handling thousands of conversations a day. Escalated leads convert into opportunities 'more than twice as often' as leads from the old form and close about five days faster, against a prior baseline of tens of thousands of monthly inbound requests and multi-day response waits. The agent is opt-in: customers choose at the start whether to talk to an agent or a sales rep.

Why: This is a vendor's own case study with no methodology, no absolute conversion numbers, and no outside verification, so treat the 2x opportunity rate and five-day-faster close as directional rather than a benchmark you can copy. The one transferable decision it does surface: the reported win came from escalating better-qualified leads, not from deflecting all of them, and the flow was deliberately opt-in — so if you're wiring an agent into an inbound funnel, measure opportunity rate and days-to-close separately instead of celebrating deflection volume.

Anthropic published a customer announcement that Barclays is expanding its use of Claude across the bank, targeting Claude Code adoption by 50% of its developer population by end of 2026 and a majority of software engineers in 2027. The post also cites Barclays' Colleague Knowledge Assistant, live since 2025 and built on a retrieval-augmented generation architecture over Claude, used by more than 16,000 colleagues supporting over 20 million UK retail customers. No measured productivity, cost, or quality results are given, and the numbers are Barclays' stated targets plus a vendor-published adoption claim.

Why: This is a vendor-published case study, not a measurement, so do not use it as evidence that AI coding tools cut delivery time. What is usable is the target shape: a large regulated bank publicly committing to 50% Claude Code adoption among developers within a year, which is a concrete benchmark to cite if you are arguing for or against an internal rollout, and a signal for founders selling AI tooling into banks or regulated enterprises that procurement conversations now assume governance and RAG-style knowledge assistants, not just model access. Nothing here changes what a Malaysian builder ships this week; there is no pricing, availability, or regional detail.

Anthropic's Claude blog published the third post in its 'human-agent teams' series, a case study of how Asana runs AI agents as teammates on its own Work Graph model, with Arnab Bose, Asana's Chief Product Officer, describing the setup. At Asana, Claude is the default AI tool connected to Google Drive, Slack, Asana, Zoom meeting recordings and Databricks reports; agents get defined roles, are assigned tasks, read and write messages, and appear in activity feeds next to humans, with extra safeguards on what they can access and share. The piece names three required capabilities: persistent memory, agents having their own credentials, and shared context.

Why: The reusable idea here is architectural, not product news: Asana did not build a separate context store for agents, it put them inside the existing task/project/owner graph, and it gives agents their own credentials rather than a shared service account. If you are wiring agents into a product or internal workflow, those two decisions are what determine whether permissions, audit trails, and 'who changed this' stay answerable. The post gives no numbers, no failure cases and no pricing, so treat it as a design pattern to compare against your own setup, not as evidence that this works at scale.

Anthropic says Claude for Government is now generally available for US federal and state agencies, running in a FedRAMP High authorized environment after a public beta that started in July. Agencies get the same capabilities as commercial customers on the commercial release cadence, plus admin controls: no seat fees, prepaid usage in fixed increments with a hard not-to-exceed cap, department-level allocation of usage to sub-agencies, identity-provider SSO, SCIM group mappings that set rate limits, dollar caps and allowed models per seat tier, and audit logs to support the agency ATO process. Claude Code CLI and Claude for Microsoft 365 are also entering early access inside the same environment.

Why: This is a US public-sector procurement story and most Malaysian builders do not have to change anything because of it. The one transferable detail is the pricing and governance shape: no seat fees, prepaid usage in fixed increments with a hard not-to-exceed cap, and per-department spend allocation with burndown alerts. If you sell AI tooling into regulated or government-adjacent buyers, that is the packaging to compare against your own seat-based plans — a fixed cap removes the buyer's main objection to usage-based AI spend.