OpenAI audit finds 30% of SWE-Bench Pro flawed, invalidating coding model evaluations

Etched ships first post-ChatGPT AI inference chip, claims 20x H100 efficiency with novel clock-alignment solution

Etched, founded by two Harvard dropouts, has achieved what industry consensus said required a decade and billions of dollars: a working specialized inference chip. The company raised $800M at a $5B valuation with $1B+ in signed contracts. Their transformer-only Sohu chip reportedly matches 20 Nvidia H100s in performance while solving a critical technical hurdle—a 50-picosecond clock alignment problem across billions of cycles that engineers initially deemed unsolvable. The team developed a novel solution in two weeks by allowing clock signals to drift naturally then freezing them at convergence. This represents the first successful silicon from a chip startup founded after ChatGPT, validating specialized inference architectures against general-purpose GPU competitors. If efficiency claims hold and deployment scales, cheaper inference could materially reduce AI inference costs across applications from medical diagnosis to tutoring, with significant implications for AI adoption economics and competitive dynamics against Nvidia's dominance. The achievement also demonstrates the technical depth and operational execution gap required to compete in AI hardware.

Modal raises $355M Series C, pivots infrastructure strategy toward AI agent workloads over developer experience

Modal's latest funding round reflects a fundamental repositioning of its cloud platform around AI agent workloads rather than human developers. The company is emphasizing that agents require tighter, more observable infrastructure primitives than traditional serverless compute, with specialized capabilities including sandboxes, GPU snapshotting, elastic inference, and networked containers designed for agent iteration loops rather than human-readable dashboards and YAML configuration. Modal has built production-grade inference capabilities including DeFlash, speculative decoding, Auto Endpoints, and elastic autoscaling across a 17-cloud capacity pool, targeting companies like Suno (audio), Runway (video), and robotics firms deploying custom models with unpredictable traffic patterns. This positions Modal as a specialized inference deployment layer rather than competing with LLM API providers or basic GPU rental services, reflecting a broader shift in infrastructure strategy as agent-based AI workloads become the primary driver of compute demand.

SpaceXAI and Cursor launch Grok 4.5: Opus-class coding model at $2/$6 pricing signals cost-based frontier model competition

SpaceXAI unveiled Grok 4.5, a 1.5T-parameter model trained in partnership with AI coding startup Cursor, positioning itself as near-Opus 4.8 capability with materially better cost efficiency and speed. Priced at $2/1M input and $6/1M output tokens versus Opus 4.8's $5/$25, Grok 4.5 ranks #4 on Artificial Analysis Intelligence Index with notably lower token consumption per task—60% fewer output tokens than Opus 4.8. Cost-per-task metrics show $0.31 for general intelligence tasks, $0.49 for banking tasks, and $2.59 for coding agent tasks, with average total tokens per coding task at 1.9M versus 7.2M for Fable 5 and 6.2M for GPT-5.5, placing Grok 4.5 on the Pareto frontier for cost/performance. The partnership represents a concrete competitive dynamic where coding-focused platforms are co-training frontier models rather than purely consuming OpenAI/Anthropic outputs. Cursor announced Grok 4.5 as "the first we've built for more than software engineering" with day-0 ecosystem support from Hermes Agent, OpenRouter, and Grok subscriptions. This launch signals a shift toward price-based competition in frontier model economics and reflects xAI's focus on the coding-agent workflow market, arriving weeks after SpaceX formally agreed to acquire Cursor in a deal valuing the startup at $60 billion.

OpenAI accelerates model cadence: GPT-5.6 launches tomorrow, GPT-6 expected within weeks on larger pretraining base

OpenAI is rolling out GPT-5.6 as the final model in the 5.x series alongside new full-duplex voice capabilities (GPT-Live), with GPT-6 slated to launch within weeks on a significantly larger pretraining base than the ~4T token foundation used for 5.5/5.6. Sources indicate OpenAI initially intended to continue with the existing base through GPT-6 but decided against it, believing the new larger base will be much better positioned to compete with Fable 5 and upcoming Fable 5.1. This accelerated cadence suggests OpenAI believes new architectural approaches are needed to maintain competitive positioning as Anthropic prepares Fable 5.1 for release in the coming weeks.

Why AI Infrastructure must evolve for Agent Experience — Akshat Bubna, Modal CTO

Latent Space published a podcast interview with Erik Bernhardsson, founder of Modal, on infrastructure design for AI agents. Main takeaways:

• Serverless inference is becoming the killer app for cloud infrastructure because AI models are stateless, use token-based pricing, and GPU scarcity makes reserved instances uneconomical
• Tight feedback loops matter equally for agents and humans; agents stall when errors are vague, configs inaccessible, or human intervention required
• Programmatic access (CLI/SDK/API) is critical for agents; UI-only features like billing, logs, and runtime environment details limit agent autonomy
• Keeping infrastructure and application logic in a single language (Python vs YAML) reduces hallucination risk and improves token efficiency for agents
• Consistent, intuitive naming across systems prevents agent errors; ambiguous names increase broken feedback loops

Separating Signal from Noise: Coding Evaluations

OpenAI published a technical report on fundamental flaws in SWE-Bench Pro, a widely-used benchmark for evaluating coding model capabilities. Main takeaways:

• OpenAI's audit found ~30% of SWE-Bench Pro's 731 public tasks contain design or contamination flaws that invalidate results, rendering the benchmark unreliable for measuring agentic coding progress.
• Four primary failure categories identified: overly strict tests enforcing specific implementations, underspecified prompts omitting hidden test requirements, low-coverage tests allowing incomplete fixes to pass, and misleading prompts pointing models toward wrong behavior.
• Frontier models showed 23.3% to 80.3% pass rate improvement over eight months on SWE-Bench Pro; OpenAI now retracts its earlier recommendation to adopt the benchmark, citing the need for benchmarks built by experienced software developers rather than sourced from open-source PR histories.
• OpenAI's quality assurance pipeline combined automated flagging, agent-based investigator review, and human annotation by five experienced engineers per task; human reviewers more frequently identified broken tasks than agents, suggesting conservative labeling of actual issues.

Beijing looks at curbing overseas access to China's top AI models

Reuters reported on Chinese government meetings with major AI companies about restricting overseas access to advanced models. Main takeaways:

• Alibaba, ByteDance, and Z.ai attended Ministry of Commerce-led discussions on limiting foreign access to China's most advanced AI models, both closed-source and open-weight versions
• Officials proposed making AI leaks or thefts punishable under national security law and discussed restricting who can fund domestic AI startups
• Chinese authorities are specifically concerned about Anthropic's Mythos model being used to exploit software vulnerabilities, mirroring U.S. national security restrictions on Chinese AI access
• Potential restrictions may apply only to future models and could follow a tiered system: basic open-source tools with simple filing, advanced tech requiring security reviews, frontier models barred from public release or restricted to domestic use