AI Agents

Nov
13
Gemini 4 Hydrogen leads Claude Sonnet 5.5 on SWE-sweep, while Sonnet leads on Terminal-Bench 4.1 and OSWorld 2.1 in Google-reported results.

Gemini 4 Hydrogen leads on SWE-sweep and trails Sonnet on terminal tasks

Gemini 4 Hydrogen has a 1M-token context window and costs $1.50 per million input tokens and $7.50 per million output tokens via the API. At its highest thinking level, it generated about four times the median output tokens in Artificial Analysis’s tests, adding to the cost per task.
4 min read
Nov
12
Pull requests queue for reviewers. A 655-firm study estimated 56.1% longer submission-to-merge time after coding-agent adoption.

Why does code review become a bottleneck when AI agents write more code?

A study of 655 firms estimated that pull requests took 56.1% longer to merge a year after coding-agent adoption. Check where your PRs wait and whether the same reviewers keep getting called in.
6 min read
Nov
10
Claude Fable 5.2 and Mythos 5.2 banner comparing Artificial Analysis index scores and cost per task across effort levels, alongside Fable 5.1.

Claude Fable 5.2 and Mythos 5.2: what do they cost and who can use them?

Token prices fell 20%, but max-effort runs cost twice as much per task on one independent index, while low effort scored above Opus 5.5 high for less. Plans, Mythos access and the API changes that can break existing agents.
9 min read
Nov
05
Diagram contrasting OpenAI keeping Git with Pando and Anthropic replacing it with Scion, both internal deployments.

Why OpenAI kept Git for its agents and Anthropic replaced it

OpenAI reports nearly 80% lower median push latency after rebuilding its Git backend. Anthropic reports growing from about 2,500 daily merges in May to over 30,000 in September. We compare those choices with new hosting and review tools, including what to measure before migrating.
9 min read
Oct
29
Gemini 4 Carbon scores 59 on Artificial Analysis Index v5.0, ahead of Claude Opus 5.5 at 58 and Argon at 50. Carbon’s rollout starts with cyber defenders; Opus and Argon are available through

Gemini 4 Carbon tops Artificial Analysis Index v5.0 as developers wait for access

Argon's introductory pricing ends November 30. Use its paid API to build an evaluation now, then compare Carbon on the same tasks when access expands.
5 min read
Oct
27
GPT-6.1 Terra and Sol compared on token prices, task costs and total output tokens.

GPT-6.1 Terra halves Sol's token prices, with smaller savings on one index

Terra targets work across documents and apps. Its cost depends on request size and reasoning effort, plus the work needed to check and repair the result.
7 min read
Oct
21
Voxtral 3 banner showing a listen–reason–speak pipeline, pricing from $0.04/min with Mistral Small 4, and enterprise phone support limited to inbound calls.

Voxtral 3 adds voice agents from $0.04/min, with inbound-only phone support

Mistral claims sub-second replies. Its own tests show a speed-accuracy tradeoff: transcription errors rise as streaming delays shrink. Voxtral 2 integrations also lose word timestamps.
4 min read
Oct
19
OpenCode Console breach banner showing saved OpenAI and Anthropic keys marked “may be exposed,” with an instruction to revoke keys at the provider.

OpenCode Console breach puts saved provider keys at risk

Users report unexpected provider bills, and Dax Raad from OpenCode says refunds are underway. He also reported further attack attempts on October 18. The steps below cover provider key replacement and spending controls while the cause and current containment status remain unclear.
3 min read
Oct
13
Meta banner comparing AA index scores at Max effort: Opus 5.5 with default fallback scores 58, Muse Wonder Full scores 56 with no public API, and Muse Wonder scores 52 with API access.

Meta's Muse Wonder Full gets close to Opus 5.5, with a subscription catch

Regular Muse Wonder costs the same $4/$20 per million input/output tokens as Opus 5.5, while scoring six points lower on an independent index. Its measured cost per benchmark task is lower, and Meta says it has fixed unusually fast quota usage on Muse Code.
3 min read
Oct
13
Banner showing a coding harness loop: prepare context, get a model response, run allowed tool calls, and record results. Pi 1.0 is identified as the case study.

What is an agent harness, and how does a coding harness work?

A runnable experiment with the Pi coding agent fails after completing an action. Running it again repeats the action. We map the execution, control, and context choices behind that result so you can test your own harness.
10 min read