AI Model Release Tracker

Anthropic just shipped Claude Opus 4.8 alongside a record $65B Series H at $965B post-money valuation.

• New capabilities: 2M token context, stronger...

Flagship Model Releases

• 🔥 Claude Opus 4.8: Anthropic released Claude Opus 4.8 with gains in honesty, coding flaw detection (4x fewer unflagged...

Nano Banana 2 (Gemini 3.1 Flash Image) and Nano Banana Pro (Gemini 3 Pro Image) are now generally available.

• Flash priced at $0.045/image, Pro at $0.134/image.
• Flash adds video input for generating context-aware images and infographics.

Technical system card analysis shows major upgrades in software engineering, agentic tool use, and knowledge work, plus reduced over-refusals and...

Anthropic is leaning hard into radical honesty with Opus 4.8, ditching the people-pleasing era for a model that pushes back on bad logic and flags its...

Anthropic's Claude Opus 4.8 launch plays on multiple fronts at once.

• IPO race: Debuts as Anthropic and OpenAI sprint toward public listings, with...

Qwen3.7-Max targets agent tasks like long-horizon coding and tool use with a 1M-token window.

• Novel training splits problems into...

No significant updates today.

No significant updates today.

EAGLE 3.1 fixes attention drift in speculative decoding by adding FC normalization and post-norm hidden-state feedback, yielding up to 2× longer...

StepAudio 2.5 Realtime swept all five April 2026 voice AI benchmarks, beating GPT Realtime 1.5 and Gemini Live with scores including 82.18 in...

SenseNova U1 delivers a native unified architecture that handles understanding, reasoning, and generation without separate vision encoders or VAEs.

-...

GPT-5.5 edges Gemini 3.5 Flash on agentic coding benchmarks while Gemini dominates tool-use and finance tasks at far lower cost.

• Coding edge:...

Anthropic’s Claude Mythos Preview scanned 1,000+ open-source projects and flagged 23,019 potential vulnerabilities, including an estimated 6,202 high-...

AlphaProof Nexus solves nine open Erdős problems and 44 OEIS conjectures via Lean-verified proofs, marking a move beyond benchmark scores toward...

WBench delivers the first unified multi-turn benchmark for interactive video world models, addressing fragmented evaluation methods.

• Spans five...

New Model Releases

• 🔥 StepAudio 2.5: StepFun released StepAudio 2.5, a unified audio-language foundation model trained on 2.2T tokens that...

Meta Llama 4 delivers native multimodal capabilities in an open-source model, processing vision, audio, and text together for cross-modal tasks like...

Anthropic’s Claude Mythos Preview scanned over 1,000 open-source projects and surfaced 23,000 potential vulnerabilities.

• Independent reviewers...

Auto Benchmark Audit (ABA) introduces an agentic pipeline that systematically scans benchmark repositories, datasets, and agent trajectories for...