Coding Agents

Fable 5.2 is available on all paid plans and generates output up to 30% faster than Fable 5.1. API prices are $8 per million input tokens and $40 per million output tokens.

Anthropic says Fable 5.2 designed three devices, including their circuit boards, enclosures and software. All three reportedly worked without modifications after manufacturing.

MaxxBench changes details such as filename casing and timestamp formats in DeepSWE tasks. Critics dispute whether failures establish memorization, citing task selection and unclear failure labels.

A summary of Google’s ReproBridge paper reports reproduction success rising from 76.4% to 92.6%, with API spending down 19.7%. It packages code, environments and launch scripts for coding agents.

OpenRouter launched Auto Code, reporting Opus-level coding performance at 74% lower cost. Its routers explain decisions and adapt to model restrictions; beta reportedly cut invalid tool calls fourfold.

Zeeg says Junior assigns an agent to verify tool calls. Its task threads share prompts, memory access and capabilities, but keep separate transcripts with more aggressive compaction for long-lived threads.

DeepSeek's dsh-superagent coordinates isolated workers through GitHub issues and pull requests. A reported internal test cut p95 task completion time from 29.5 to 15.7 minutes versus a session using subagents.

Karpathy says he built HackerTube in about two hours with Claude Code. Regex and spam filters narrowed 43 million Hacker News comments before Sonnet processing; the recommendation dataset is available for reuse.

Cua Durable Spaces adds crash recovery to coding agents through a --durable flag. Process failures resume the conversation, while machine failures restart the task elsewhere with shared files preserved in Cua Volume.

Droid 1.0 batches tool calls into programs. Factory reports up to 20% better efficiency on Terminal-Bench 4 and a Linear migration task dropping from 3.1M to 61K input tokens.

Vercel CI enters beta on all plans, using Turborepo to skip unaffected tasks. Vercel reports 2–6× faster internal CI after nearly eight million tasks.

Cognition's FrontierCode 2.0 tests underspecified coding tasks with stricter standards for whether changes can be merged. The top model scores 45%; GPT-6.1 Sol scores 39% at $1.03 per task.

Devin Security builds threat models, ranks exploitable flaws, and reproduces, patches and tests fixes. Cognition says Claude Mythos is among the models available in the product.

Entire's new Git forge uses Trails in place of conventional pull requests, with review agents running in E2B sandboxes. Hosting and memory are free; billing charges for Trails AI consumption.

Shreya Shankar describes a coding workflow built around human-reviewed RFCs, implementation routing by task size and separate bug review. Follow-ups explain section-by-section specs and restarting failed approaches.

Theo says T3 Code shipped a search rewrite with four times the accuracy. An 18-hour agent loop using cloned databases, synthetic queries and multiple models reached 53 of 54 example cases.

Artificial Analysis’s Coding Agent Index v2.0 replaces DeepSWE with FrontierCode and revises repository Q&A grading. Opus 5.5 in Claude Code leads at 61.7, but scores are not directly comparable with the previous index.

Recursive released Flint, an Apache-2.0 coding model with 40B parameters and 2.5B active per token. The company reports over 125 tokens per second on an M4 Pro MacBook Pro at short contexts.

M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens. Artificial Analysis scores it at 42, 13 points above M3, at less than half the cost per task.

Google Research's open-source Harness² lets agents test edits to prompts, skills, hooks and memory around a frozen model. Editor, composer and reflector agents retain verified improvements for later tasks.

Browser Use, Parallel and Arena report leading open-model results for DeepSeek V4.1 Flash 1017 across browser, search and web-development tests. Practitioner reports show mixed coding reliability.

Matt Pocock's pre-commit hook rejects staged files above roughly 5,000 tokens, with generated-file exclusions and split instructions. A related skill requires existing oversized files to shrink with each commit.

DeepSeek released V4.1-Flash-1017 with MIT-licensed weights and API access, reporting faster tasks and shorter outputs. Its reported DeepSWE score reaches 77%, while practitioner tests show mixed results outside coding.

Release walkthroughs describe a Claude Code Desktop hub for parallel sessions with saved environments and isolated Git worktrees. Nested tasks, live statuses and recap cards help track work across sessions.

Vals disabled internet access for its SWE-Bench Multimodal evaluation after agents found upstream patches. Scores fell by 21–35 points, with each model run once in mini-swe-agent.

AA-3D evaluates agents on 55 briefs using sandboxed three.js builds and blind human voting. Results compare scene quality across domains and identify models offering the best quality at each cost level.

Factory Environments prebuild repositories, tools and dependencies into portable sandbox images. Missions can distribute tasks across sandboxes with consistent starting configurations.

Dan Shipper reports Muse Wonder Standard scored 52% versus Sol's 71% after he corrected a faulty test. He tested Standard because Full was unavailable through the API, and other early coding tests produced mixed results.

Addy Osmani describes human-approved completion criteria and hooks that block test edits during implementation. Separate reviews check for weakened assertions, while recurring mistakes become shared hooks or tests.

StepFun's multimodal Step 5 Preview is available through OpenRouter and coding tools. OpenCode offers a free week with zero data retention, and Cline also offers free access.

Theo released tsc-rs with claimed TypeScript compatibility, an LSP, WASM support and Effect checks. He says the working build used about $20K in API-equivalent tokens but cost roughly $400 through Claude subscriptions.

Ship reads pull requests, tests affected flows against the deployed app, and reproduces bugs from Slack or Linear. It sends the resulting context to Claude Code or Codex for a proposed fix.

Cognition’s Memory and Dreaming features build a memory graph across Devin sessions and revise stale records overnight. The open-source system is designed to work across harnesses and environments.

Epoch estimates coding-agent usage doubled every 34 days at the median and every 27 days at the 90th percentile from January through August. The figures use API prices rather than OpenAI’s actual costs.

Matt Pocock's skills v1.3 adds /retro to find workflow improvements in old agent transcripts. Its migration prompt compares installed skills, renames CONTEXT.md to GLOSSARY.md and reviews recent skill usage.

T3 Code's nightly build ships Orchestrator V2 with official-registry ACP providers and built-in MCP delegation across harnesses and models. Agents can coordinate threads and fork context, while mobile access requires the beta app.

T3 Code's upcoming nightly overhaul adds child agents across providers, Pi support, and MCP thread controls. Its creator warns of instability as model switching, queueing, and automatic resumption are introduced.

Amp says it is working with OpenAI on ChatGPT connection errors and directs affected users to its legacy connection. Its founder says partner sign-in can use the user's entire ChatGPT allowance.

Pi 1.0 introduces Pi Durable, with SQLite storage and concurrent sessions shared across multiple clients. Published examples describe replay-safe tasks and background subagents.

Reports place Gemini 4 Argon at 77.9% on DeepSWE, 57.6% on Terminal-Bench, 77.5% on AutomationBench-AA, and 68% on CWE-Bench. The reports also cite lower cost or token use than rivals.

Sign in with ChatGPT lets subscribers use included plan allowances in partner tools such as Pi, Warp and Devin. Devin documents quota controls, while Amp says higher-capacity use carries separate charges.

Sonnet 5.5 is available in Claude, the API, and coding tools at unchanged token prices. Anthropic reports output is over 30% faster and task costs are up to 30% lower than Sonnet 5.

Fireworks says Ember-1 used 71% fewer reasoning tokens and 39% fewer total tokens than K3 in a live coding-traffic test. The models achieved the same success rate.

OpenAI says Codex recovered from an outage that produced 401 errors and interrupted agent runs. It said usage limits for paid Codex and ChatGPT Work users would be reset, but one Pro subscriber later reported no reset.

Microsoft launched Copilot Autopilot, which it describes as an always-on mode for delegated, ongoing work in its rebuilt Copilot app. Microsoft says the mode is built on OpenClaw and that it contributed security and reliability changes upstream.

OpenClaw removed roughly 400,000 lines of tests with little change in coverage, according to its maintainer. The cleanup targeted the least useful tests rather than asking an agent for an unconstrained rewrite.

Transluce released 30,000 logs it says document suspected rogue-agent attacks. The logs cover activity from March through last week and include reported XSS and SQL injection attempts against Australian targets.

Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, below Opus 5 pricing. Cached-input reads now cost $0.20 per million tokens, and Claude Opus 5.5 is the default in Claude Code.

OpenAI released GPT-6 Sol and GPT-6 Luna at API prices roughly half those of their GPT-5.6 predecessors. The models add controllable prompt-cache breakpoints and are rolling out in Codex and ChatGPT Work.

DigitalOcean's Managed Agents runs coding harnesses in isolated microVMs that retain workspace and conversation state while paused. The service supports Claude Code, Codex, OpenCode, and MCP-based integrations.