LLM Benchmark Watch

Frontier Model Releases

• 🔥 Claude Opus 4.8: Anthropic released Claude Opus 4.8 with 69.2% on SWE-Bench Pro, 4x fewer unflagged code flaws than...

No significant updates today.

No significant updates today.

Model Release Plans

• 🔥 Anthropic Mythos Public Release: Anthropic plans to release Mythos-class models publicly once stronger safeguards are...

Bedrock shines for AWS-native, regulated workloads by unifying IAM, VPC, and compliance under one bill.

• Latest models: Direct provider APIs win with...

• Anthropic plans to release Mythos-class models publicly once stronger safeguards exist, after restricting its elite bug-finder due to misuse risks
  -...

• First large-scale formal theorem proving benchmark for LLM agents marks a specialized milestone in math research eval
• Broader 2026 research...

Enterprises face escalating AI debt that traditional technical debt frameworks cannot address.

• Prompt debt emerges from undocumented tweaks,...

Two releases highlight the push toward leaner, more reproducible LLM infrastructure.

• OSCAR from Together AI delivers attention-aware 2-bit KV cache...

• MolmoAct2 from Allen Institute for AI beats Physical Intelligence's π0.5 on real-world and simulation benchmarks while running up to 37x faster.
  -...

Gemini Audio identifies speakers and detects emotional tone in recordings, moving past basic transcripts.

• Exports structured insights as JSON for...

Researchers argue AI agents must be secured as untrusted systems, not by hardening underlying models.

• Five principles from systems security—least...

An LLM trained on an IBM quantum computer correctly answered questions the base model failed on.

Pricing Updates

• 🔥 DeepSeek V4-Pro Discount: DeepSeek made its 75% price cut on V4-Pro permanent, setting output tokens at $0.87 per million and...

• V4-Pro pricing now permanent: Output tokens at $0.87 per million, 29x cheaper than Claude Opus and 34x below GPT-5.5.
• Agent use cases unlocked:...

Multiple players are pushing agentic coding beyond traditional IDEs.

• Google Antigravity 2.0 launches as a standalone app with Gemini 3.5 Flash (289...

Anthropic's Project Glasswing shifted posture in one month: from "we will not release" Mythos Preview to "we look forward to releasing" once stronger...

A novel Lean-based framework evaluates LLMs on open research-level math proofs, solving 9 of 353 Erdős problems—including some unsolved since 1970. By...

Cerebras's WSE-3 packs 4 trillion transistors on one dinner-plate-sized chip to eliminate the inter-GPU communication that dominates large-scale training, offering a direct path to faster iteration and lower costs when scaling frontier models.

Prompt engineering is shifting from ad-hoc practice to structured discipline amid rising AI agents.

• Tooling emerges: Langfuse enables end-to-end...