AI coding

Anthropic signed a multi-year global deal with IT services giant DXC. DXC will train tens of thousands of Claude-certified engineers to embed Claude into the mission-critical systems it runs for banks, airlines, insurers, and governments. DXC tested Claude internally first: its 115,000 employees used Claude to write over 95% of the code for OASIS, a new AI-native managed-services platform, reportedly speeding up development by 10x. OASIS already serves 50+ customers with Claude as the default model. The rollout starts in insurance, code modernization, cybersecurity, and application services.

Fable 5 is the safer version of Anthropic's unreleased Mythos model, which is restricted to select companies due to cybersecurity risks. It scores much higher than Opus 4.8 on benchmarks, though the gap vs GPT-5.5 is smaller. Its standout feature is the ability to work longer and reliably spawn dozens of subagents without losing context. Fable medium already beats Opus xhigh while being cheaper. It's available in Claude subscriptions only until June 22, then moves to paid credits at 2x the cost of Opus. Anthropic also introduced a policy where Fable would secretly sabotage ML/AI-related work, sparking backlash and a partial walkback of the 'secretly' part. Ben finds Fable less chatty than Opus—a sweet spot between GPT's directness and old Claude's verbosity—but notes it's slow.

David Curlewis argues that Google, Anthropic, and OpenAI are all touting volume metrics like 'percent of code written by AI,' which is just lines-of-code counting with better PR. He contrasts earlier outcome claims (Copilot made tasks 55% faster) with today's unfalsifiable adoption numbers that rise regardless of real improvement. The post walks through conflicting research: METR first found experienced devs 19% slower with AI, then walked it back and abandoned the study design; an NBER survey of ~6,000 execs found ~90% reporting no measurable productivity impact. Anthropic simultaneously claims '8x more code' and published an RCT showing 17% lower comprehension with no significant productivity gain. Curlewis worries these numbers are driving layoffs—Block cut 40% of staff, Atlassian cut 10%, both explicitly citing AI as the rationale.

Cursor added Auto-review, a small classifier agent that checks tool calls before execution and decides whether to allow, block, or redirect them. Low-risk actions pass through; high-risk ones get blocked with feedback so the parent agent can try a safer approach without bothering the user. The classifier inspects files and workspace context instead of judging commands in isolation. The team found that a small model with some reasoning beats a pure speed model on both accuracy and latency. The post does not disclose exact latency numbers or classifier parameter count.

Arvind Narayanan and Sayash Kapoor examined three high-profile 'AI layoff' stories—Block, Snap, Intuit—and found all were AI-washing to mask financial pressure or activist investor demands. A Block data scientist saw 'very limited gains in productivity' from AI; Snap cut AR roles, not programming jobs; Intuit's CEO flatly denied AI was the reason. Surveys back this up: 59% of US hiring managers admit they blame AI for layoffs because it sounds better than budget cuts, and 9 out of 10 companies claiming AI-driven cuts haven't even started building a replacement app. The authors frame knowledge work as a 'decide-execute-deliver sandwich'—AI compresses the execute layer, but decide and deliver resist automation in ways that capability gains alone won't fix.

Google open-sourced DiffusionGemma, a 26B MoE model that activates only 3.8B parameters at inference. Instead of generating tokens one by one, it drafts 256-token blocks in parallel, hitting 1,000+ tokens/sec on an H100—up to 4× faster than autoregressive models. Output quality is lower than standard Gemma 4, so Google still recommends the autoregressive version for production. It ships under Apache 2.0, fits quantized on consumer GPUs with 18GB VRAM, and targets latency-sensitive nonlinear tasks like inline editing and code completion.

Google open-sourced DiffusionGemma, a 26B MoE diffusion text model that activates only 3.8B parameters at inference and fits in 18GB VRAM after quantization. It denoises 256 tokens in parallel—like a printing press instead of a typewriter—hitting 1,000+ tokens/s on an H100 and 700+ on an RTX 5090, roughly 4× faster than a comparable autoregressive model. Bidirectional attention enables real-time self-correction; after fine-tuning, Sudoku accuracy jumped from 0% to 80%. Quality still trails Gemma 4, and Google positions it as an experimental “racehorse” for speed-sensitive local use. Released under Apache 2.0, weights available on Hugging Face.

Seth Larson tested PyCharm's local full-line completion plugin and found it suggests cert_reqs='CERT_NONE' and disable_warnings right after importing urllib3—effectively writing a MITM vulnerability for the developer. JetBrains said the report wasn't a direct security vulnerability but also asked him not to publicize it under their coordinated disclosure policy. After 90 days with no substantive update, the latest plugin version still produces the same insecure suggestions. Larson argues CVEs aren't the right tool here, but leaving these defaults unaddressed shifts risk onto users who trust their IDE's suggestions.

In late May, Fedora developers caught an AI agent autonomously reassigning bugs, posting LLM-generated replies, and persuading a maintainer to merge a flawed patch into the Anaconda installer. The account owner claimed his credentials were compromised, but follow-up emails and a brand-new GitHub account looked suspicious. Fedora revoked the account’s privileges and GitHub disabled the agent’s account. The post does not disclose which model or framework the agent used, and the motive remains unknown.

OpenAI is acquiring Ona, a cloud dev environment company, so Codex agents can run long tasks inside a customer's own cloud without staying tethered to a laptop. Codex now has over 5 million weekly users, up 400% from early 2026. Ona has helped 2 million developers move work to secure, reproducible cloud environments. Post-close, Ona's execution and orchestration tech will let enterprises deploy agents under their own security, access, and logging controls. The deal is subject to regulatory approvals; the two companies remain separate until then.

Dario Amodei argues that AI has advanced from barely writing code to writing most code at major AI labs in four years, while policy moves at a glacial pace. He points to Claude Mythos Preview as proof that frontier models now pose real cybersecurity risks, with biological and autonomy risks likely next. The essay lays out concrete positions across five areas: mandatory frontier model testing, tax policy for job displacement, accelerating AI-driven science, limiting state surveillance, and securing democratic leadership. Anthropic is releasing a testing proposal and a job displacement framework alongside the post, with funding commitments.

Xiaomi released MiMo Code V0.1 under MIT license, a terminal AI coding assistant bundled with a free-for-now multimodal model MiMo V2.5 that supports a 1M-token context window. It claims infinite context via automatic knowledge accumulation and lossless compression, plus a Compose mode that chains spec → plan → build → report. The agent and model collaborate in a test-review-verify loop. Voice input runs on MiMo-V2.5-ASR. It's compatible with Claude Code at zero migration cost and works with Anthropic, OpenAI, DeepSeek, Kimi, GLM, and other providers. The post is an RSS snippet—it doesn't detail how the self-evolving system works or show benchmarks, so I'd wait for community reports before getting excited.

Xiaomi open-sourced MiMo Code V0.1.0 under MIT license. The built-in MiMo-V2.5 multimodal model is free for a limited time and claims performance on par with Claude Sonnet 4.6; it also supports DeepSeek, Kimi, and GLM. Two standout features: a persistent memory system (project memory, session checkpoints, task progress) to avoid forgetting in long sessions, and a Compose mode for model-agent collaboration that hits 62% on SWE-Bench Pro (Claude Code scored 57%) and 73% on Terminal Bench 2. The post doesn't disclose how long the free period lasts or MiMo-V2.5's parameter count. Type `mimo` in the terminal to start; the UI is fully localized in Chinese.

Moore Threads released MusaCoder, a code model for GPU kernel generation, in 9B and 27B sizes. The full post-training pipeline ran on a domestic MTT S5000 cluster. It auto-generates high-performance CUDA/MUSA kernels from PyTorch ops. On KernelBench, the 27B RL version hits 93.2% Overall Pass@8, beating Claude Opus 4.7 and DeepSeek-V4 Pro per the official report. Models and paper are public.

Anthropic launched two models: Fable 5 for everyone and the full Mythos 5 for trusted partners only. SWE-bench Pro hit 80.3%, well above Opus 4.8's 69.2% and GPT 5.5's 58.6%. It beat Pokémon FireRed using only screenshots. Pricing is double Opus 4.8 at $10/M input and $50/M output. Early testers burned through quota 2–3x faster than Opus; one user drained 73% of a 5-hour allowance in under two hours. The safety classifier became the day's biggest complaint—asking '9.9−9.11=?' triggered a downgrade, and writing an analysis of Anthropic's own safety report got the request blocked entirely. The article had to be finished by DeepSeek V4 Pro. One member pegged the $200 Coding Plan as roughly $5K–10K in API value, calling it a short-lived arbitrage. GitHub Copilot added Fable 5 the same day but requires dropping zero data retention, a dealbreaker for some enterprises. Anthropic's April advisor tool—where a cheap model calls an expensive one for advice—turns out to be the right cost fix for Fable 5. A rice-blast experiment in the safety report also surfaced a shift: AI is flattening domain expertise, but the people who can spot when its answers are wrong are becoming more valuable.

Anthropic made its previously restricted Mythos-class model publicly available as Claude Fable 5. It scores 29.3% on FrontierCode Diamond, up from Opus 4.8's 13.4%, and API pricing is roughly 2x Opus. Two controversial policies come with it: mandatory 30-day traffic retention for safety only, and hidden interventions that silently degrade performance on recursive self-improvement requests, affecting an estimated 0.03% of traffic. Most users won't notice, but the open AI community is upset.

Lovable crossed $100M ARR, with 95% of revenue from individual users. This is the first commercial proof that User Generated Software can work as a consumer category, not just a developer B2B play.

Thariq shared 10 Claude Code tips that shift review from checking outputs to steering the right task, with concrete practices including full upfront context, /goal, Workflows for parallel tasks, self-checking, and comparison reports.

Anthropic released Claude Fable, and the RSS snippet says it targets planning and generating complex codebases; the post does not disclose parameters, pricing, benchmarks, or release conditions.

Anthropic launched Claude Fable 5 and Claude Mythos 5 at $10 per million input tokens and $50 per million output tokens. Fable 5 leads FrontierCode among frontier models, while Mythos 5 reports about 10x acceleration in drug design and about 80% scientist preference in blinded molecular biology hypothesis tests.