A benchmark of 153 autonomous runs across 18 frontier models on the nanoGPT optimizer speedrun shows Fable 5, run via claude-code at high effort, achieved the best validated result of 2,726 tokens wit…
Anthropic's Claude prompt caching silently fails in agent loops when the conversation exceeds 20 content blocks between cache breakpoints, causing cache reads to drop to zero and triggering full-prefi…
Aikido Security burned 11.7 billion tokens to benchmark 10 AI models on rediscovering 32 fresh vulnerabilities, finding that DeepSeek V4 Pro 0813 found the most (28 of 32 across three runs) and that o…
Ben Hoyt, a manager at Canonical, spent his two-week mid-year break upgrading his side project Gifty Weddings from a wedding gift registry to a wedding website builder, using AI tools Claude Code with…
A user's informal test comparing ChatGPT and Claude responses to a humorous gif found ChatGPT's reply more concise, while Claude's response was criticized as verbose and flowery. The user, who calls t…
Dromeas, a code review tool, compared its LLM council-based review against Claude Code's ultrareview on a large, real pull request from the open-source project openclaw. Dromeas's three-model council …
Anthropic's Opus 4.8 and Opus 5 language models are generating confusing or invented terminology, forcing developers to spend extra time and tokens on cleanup, with token costs up to two times higher,…
Claude Code version 2.1.237 introduces a built-in Concise output style that trims verbose responses, replacing the need for custom ASD-STE100-based styles. In tests, the Concise style reduced a 663-wo…
A developer revived a Microsoft Band 2 that was stuck in setup mode by using Claude Code on Opus 5 and the open-source msband-py library, which speaks the Band's USB protocol without a phone. The Band…
Rémy Salim's team at Quarkslab spent two weeks testing whether LLM-assisted reverse engineering defeats obfuscation, handing Claude Code agents a series of hardened AArch64 binaries with one prompt: r…
Anthropic's Opus 5 model improves self-verification but shifts the bottleneck to code review, as developers face larger diffs and longer review times. Telemetry from Faros AI and LinearB shows median …
GLM-5.3 tied the leading open weights score with 60 on the Artificial Analysis Intelligence Index, matching Kimi K3, and posted a 246-point jump in agentic Elo from 1524 to 1770, second only to Opus 5…
A benchmark comparing 8 AI models on ASCII-art generation across 3 prompts found Kimi K3 fastest at 14 seconds for the first prompt, while Fable 5 took 353 seconds and cost $1.229, and Opus 5 complete…
Anthropic's Claude Code and Cursor offer different approaches to AI coding, with Claude Code excelling at autonomous task completion while Cursor provides faster responses but requires more user overs…
Meta launched Muse Code on August 5, an AI coding agent built on Muse Spark 1.2 that uses parallel sub-agents in isolated git worktrees, a feature not offered by Claude Code or Codex CLI. However, the…
A developer discovered that Claude Code subagents pinned to run on the cheaper Sonnet model were silently falling back to the session model, Fable, due to a frontmatter pin bug, causing unexpected cos…
A new public experiment by Prime Intellect ran 153 autonomous runs on the nanoGPT optimizer speedrun across 18 frontier models, finding that Claude Fable 5 and Opus 5 dramatically outperformed others,…
Ctok, an unofficial open-source library, reconstructs Anthropic's Claude tokenizer offline, reporting exact token counts for 1,664,940 v3 and 1,722,961 v4.7 texts with zero under-counts. The library s…
Benchling released BenchBench-Protocol, a benchmark built from thousands of real-world experiments, showing that Anthropic's Opus 5 leads at 59.2%, followed by OpenAI's GPT 5.6 at 47.1%, and open-sour…
Anthropic apologized after secretly routing paying Fable 5 customers to the cheaper Opus 4.8 model while billing them for the premium tier, a practice that sparked developer backlash and was fixed onl…