Grok 4.5 vs GPT‑5.6: cheap coding, cheaper IQ

SpaceXAI launched Grok 4.5, a frontier model tuned for coding, agentic workflows, and complex knowledge work. Priced at $2 per million input tokens, it undercuts competitors while delivering high‑level reasoning and tool integration. The model is already available via the xAI API and in Cursor.

OpenAI today released the GPT‑5.6 family, Sol, Terra, and Luna, claiming state‑of‑the‑art performance on coding, knowledge work, cybersecurity, and science at dramatically lower token cost. In the Agents’ Last Exam benchmark Sol hits 53.6, outpacing Claude Fable 5 by over 13 points while cutting estimated spend to a quarter. The rollout also adds the ‘ultra’ setting for multi‑agent parallelism and the strongest safety layers yet.

Mistral unveiled Robostral Navigate, an 8 billion‑parameter model that lets robots move autonomously using only a single RGB camera. It hits 76.6% success on the unseen R2R‑CE benchmark, outpacing the best multi‑sensor systems by up to 4.5 points, and runs on wheeled, legged and flying platforms.

OpenAI’s audit of SWE‑Bench Pro uncovered that over a third of its 731 public tasks are broken, tests are too strict, prompts underspecified, or coverage low. This calls into question recent claims of massive progress in AI coding, and suggests the community must rethink benchmark design before trusting model capabilities.

Anthropic’s new Gradient‑Routed Auxiliary Modules (GRAM) add removable neural compartments for specific dual‑use domains, letting a single model be filtered on‑the‑fly without retraining separate versions. If it works, developers could safely expose biotech or security knowledge only to vetted users while keeping the base model’s performance intact.

The paper shows how to locate LLM 'personas' in a latent weight space using OCEAN traits, then steer them with low‑rank adapters that boost or suppress Openness, Conscientiousness, Extraversion, Agreeableness, or Neuroticism. This enables controllable personality shifts while preserving core capabilities, opening new safety‑focused editing tools.

Using Anthropic's J‑lens, the author probes LLaVA‑1.5‑7B’s internal visual workspace and finds the model silently knows an object is absent, yet answers “yes” to yes/no questions. Switching to a forced‑choice format restores accuracy, suggesting hallucinations stem from question framing rather than missing knowledge.

The paper proposes ‘overthinking’, adding a scaled difference between a reasoning‑distilled model and a standard model to boost reasoning‑related weights at inference. Experiments on 2B, 32B language models show this amplification surfaces hidden information and misalignments up to ten times more often than plain reasoning prompts.

Researchers introduce Gradient‑Routed Auxiliary Modules (GRAM), which isolate risky knowledge in dedicated model components that can be switched on or off. A single model trained with GRAM can imitate multiple versions filtered for different dangerous data, cutting training costs while preserving performance on permitted tasks.

Google Cloud now offers AlphaEvolve, a Gemini‑powered AI agent that refines existing algorithms into faster, human‑readable code. It tackles hard problems in chip design, logistics routing, and medical research, already proving its mettle with early adopters like BASF and JetBrains.

The AI Futures Project released “AI 2040: Plan A,” a detailed scenario that pushes superintelligence development to 2040 through global research transparency and a mutually‑assured compute‑destruction pact. The plan gives policymakers a concrete, step‑by‑step roadmap to curb an AI race and reduce existential risk.