Gemini lets you edit the file inside the conversation that created it
PDF editing reaches web and mobile, while the other listed document formats remain web-only in this release.
The signal
All stories
Gemini 4 Carbon starts with security partners, not a general release
Google’s new security model pairs strong reported benchmark results with restricted access. The evaluation conditions matt...
Strava connects to Gemini, but live training data still needs careful interpretation
The optional connection is read-only and requires a Strava subscription. Better questions matter more than a confident tra...
Nvidia’s Nemotron 3.5 Transcribe lets the instruction shape the transcript
Open weights, speaker labels and word timestamps broaden the workflow. The fastest published runtime is not yet in the ope...
YOUR HARDWARE.
YOUR AI.
Explore GPUs, local AI systems and model memory requirements.
Capability, not hype.
Explore comparisonsMirrorCode
Whole-program reimplementation with hidden tests; ML + Private, two languages. Budget: seven days and ten billion tokens per attempt.
| Model | Configuration | Result | Test date |
|---|---|---|---|
| Claude Sonnet 5.5 | Claude Sonnet 5.5 (max) | 97.8* % | 2026-10-05 |
| Claude Opus 5 | Claude Opus 5 (max) | 87.8* % | 2026-10-20 |
| Claude Opus 5.5 | Claude Opus 5.5 (high) | 78.9* % | 2026-09-22 |
| Claude Haiku 5.5 | Claude Haiku 5.5 (max) | 78.9* % | 2026-10-21 |
| Claude Fable 5.1 | Claude Fable 5.1 (high) | 73.3* % | 2026-09-10 |
| Claude Opus 4.8 | Claude Opus 4.8 | 65.6* % | 2026-10-21 |
A closer look
More from AI
Genie 4 adds sound to interactive worlds, but consistency remains the real test
Google’s research preview connects audiovisual simulation with an observing agent. Its limits matter more than a convincin...
A fictional £350 refund shows how an AI agent can pursue the wrong kind of success
Vals’ experiment pressures agents after a missed deadline. The useful evidence is in their decisions, not a six-model lead...
RAMDeck can spread a model across computers. It cannot turn their memory into one fast pool
Layer splitting can make a model fit. Network latency, unequal devices and benchmark choices determine whether the result...
* Provisional data