Research
Papers, experiments and benchmarks on what Claude models can and cannot do.
-
Research Claude Code ecosystem tracker: plugins, skills, MCP stats
A weekly count of the Claude Code ecosystem on GitHub: 16,586 plugin repos, 16,784 skill repos, 7,725 MCP servers and 1,004 mods, with the fastest-growing and newest repos and links to our reviews.
-
Research Gemini 4 Neon vs Claude Opus 5.5 and Sonnet 5.5: price, per-task cost and when to switch
Google released Gemini 4 Neon on November 18 at $3 input and $15 output per million tokens, priced between Claude Sonnet 5.5 and Opus 5.5. On Artificial Analysis it scores 61 at xhigh for $6.44 per task. Here is how it compares, and why the answer changes on December 1.
-
Research Tencent Hy4 flash: price, coding benchmarks and how it compares to Claude Haiku 5.5
Tencent's open-weight Hy4 flash costs $0.070 input and $0.281 output per million tokens, and OpenRouter currently charges half that. Tencent's own tests put it close to Claude Haiku 5.5 on office tasks but far behind DeepSeek V4.1 Flash on terminal coding.
-
Research GPT-6.2 benchmarks vs Claude Fable 5.2: what independent evaluators found
Independent tests of GPT-6.2 Astra and Sol from Artificial Analysis, Vals AI and Epoch AI: Claude Fable 5.2 leads the broad indexes, while GPT-6.2 costs a fraction per task.
-
Research Gemini 4 Hydrogen vs Claude Sonnet 5.5 for coding: price, speed and benchmarks
Gemini 4 Hydrogen costs $1.50/$7.50 per million tokens against $2/$10 for Claude Sonnet 5.5. Google's own table has each model winning different coding tests, and an independent index puts them one point apart. Here are the published numbers, with sources and dates.
-
Research Gemini 4 Carbon vs Claude Opus 5.5 and Fable 5.2: benchmarks and pricing compared
Gemini 4 Carbon costs half of Claude Opus 5.5's per-token price until November 30, then exactly the same. Google's Carbon table and Anthropic's Fable 5.2 table each include the other's models, and they disagree on several rows. Here are the published numbers, with sources and dates.
-
Research FrontierHarness Eval v1.1: with open models, Claude Code was among the costliest harnesses per solved task
Runta and Baseten ran 22 configurations of 18 coding-agent harnesses with five open models over 62 terminal tasks. The best-value harness changed with the model, and Claude Code was the second most expensive per solved task. Closed models were not tested.
-
Research Claude Code on GitHub by the numbers: 121,725 repos, 24,336 of them new in 30 days (November 2026)
121,725 public GitHub repositories carry the claude-code topic as of November 17, 2026, and 24,336 were created in the last 30 days. Counts by category, stars, language and month, the top repos, the week's risers and the queries behind the counts.
-
Research ystack says its Claude Code mods cut usage 78%. Its own write-up calls that the upper end
ystack, a new bundle of 35 Claude Code mods, says it cut usage about 78% per hour of work. The figure comes from one heavy user's logs, and the author treats 39% per API call as the firm number.
-
Research Researcher says released weights show six Chinese models were started on Nvidia Blackwell GPUs
Will DePue's checkpoint-forensics report says the starting values left in published weights reveal which GPU created them, and that six models from DeepSeek, Moonshot, MiniMax and Ant Group match Nvidia Blackwell. How the method works, and what it cannot show.
-
Research Fable 5.2 designs three working circuit boards
Working with PCB maker Diode, Anthropic had Claude Fable 5.2 design a USB speakerphone, a voice-assistant hub and a handheld virtual pet, from schematic to enclosure and firmware. The boards worked without circuit fixes. What it did and where it fell short.
-
Research Fable 5.2 formalizes Fermat's Last Theorem in 30 hours
Anthropic says 1,000 agents on an early Fable 5.2 reproduced a machine-checked Lean proof of Fermat's Last Theorem in 28.7 hours, against 11 days the first time. What the agents were given, how they organized themselves, and the safety setup.
-
Research Apple paper: Claude Opus 5.5 runs a robot arm by writing its own code
Apple researchers' AMP1 harness lets a general-purpose model drive a robot arm by writing Python. With Claude Opus 5.5 it succeeded in 104 of 108 real-world trials and 82.8% of RoboLab-120 simulations (default instructions), at about $1.39 per simulated episode. How it works and the code.
-
Research Study: exact values cost Opus 5.5 a third of its auction gains
Does giving AI agents more information help them strike better deals? A paper from Stanford, the University of Maryland and All Purpose AI tested ten models trading, bargaining and bidding. The answer depends on the model, the market and how long they talk.
-
Research How AI is starting to change scientific modeling
In a guest essay on Anthropic's Science Blog, Tübingen researchers Jakob Macke, Daniel Gedon and Stefan Wahl split scientific modeling into three jobs and argue AI should return a range of plausible models, not one best fit. A summary, plus how to try simulation-based inference.
-
Research Anthropic used Claude to map the first prokaryotic vault particle in 3D
Anthropic's biology lab directed Claude through finding, isolating and mapping the first 3D structure of a vault particle from prokaryotes: 96 copies of a phage protein, about 90 minutes from cells to microscope grid, a 34-hour first model, and the mistakes along the way.