OpenAI cuts GPT-5.6 prices as Google unveils Gemini Robotics 2

OpenAI cut API prices for GPT-5.6 Luna by 80% and Terra by 20% on July 30, 2026. The new rates took effect immediately. Sol’s price is unchanged, but a faster API option is now available for that model.¹ OpenAI says the lower prices also affect how Luna and Terra usage is counted in Codex and ChatGPT Work.²

The cuts reduce per-token costs for the family’s two less expensive models. Amazon Bedrock is also automatically updating its on-demand prices to match.³¹ OpenAI has not explained exactly how the usage calculation changes in Codex and ChatGPT Work.² Companies should therefore assess API spending separately from usage limits before projecting savings across existing workflows.

Google DeepMind introduced three robotics models on July 30, 2026. Gemini Robotics 2 is designed to control humanoid robots from feet to fingertips, extending beyond the upper-body focus of earlier systems. Gemini Robotics ER 2 handles video understanding, task planning, and collaboration among robots. Gemini Robotics On-Device 2 is designed to run locally on robot hardware. Access to the two robot-control models is initially limited.¹²

Gemini Robotics ER 2 is available through the Gemini API and Google AI Studio, with private preview access on the Gemini Enterprise Agent Platform.²³ That difference lets companies test planning capabilities before they can test direct full-body robot control. How reliably the control model performs beyond demonstrated tasks, and when access will widen, remain open questions.¹

On July 28, 2026, Anthropic presented two cryptanalysis results aided by Claude Mythos Preview.¹ The model identified a previously unknown weakness in HAWK, a post-quantum digital signature candidate under NIST review.¹ For the small HAWK-256 parameter set, Anthropic estimates that the work needed for full key recovery falls from 2⁶⁴ to 2³⁸ operations.¹

The model also found a faster attack against a reduced-round research version of AES-128.¹ HAWK is not yet an adopted standard and had already undergone extensive expert review.¹ The finding illustrates how AI can assist evaluations of cryptographic candidates.¹ Neither result demonstrates an attack on encryption currently used in production: the AES result does not apply to the standard cipher, and the HAWK finding has not been shown to extend to other parameter sets.¹

NVIDIA and Safe Superintelligence (SSI) announced a long-term strategic partnership on July 27, 2026. NVIDIA is investing in the AI research lab, which will also gain access to its upcoming Vera Rubin platform.¹ Reuters put the investment at $5 billion, citing a person familiar with the matter. NVIDIA did not disclose an amount in its official announcement.¹²

SSI plans to expand its research with the funding and compute access. The companies say the agreement will enable SSI to increase its compute capacity by an order of magnitude. They also plan to collaborate on current and future NVIDIA platforms.¹ SSI was founded by Ilya Sutskever, OpenAI’s former chief scientist.³ The agreement concerns research infrastructure, not the release of a new model.¹

The Model Context Protocol (MCP) published a new specification on July 28, 2026.¹ Its core no longer uses protocol-level sessions: the initialize and initialized handshake and the Mcp-Session-Id header have been removed.² Version and capability information now travels in each request’s _meta fields.² Anthropic is rolling out support across Claude products.³

The change makes MCP servers easier to run on standard HTTP infrastructure, including serverless and edge deployments.¹³ Operators can reduce their reliance on sticky sessions and shared session stores.¹ A stateless protocol does not make every application stateless: workflows that span multiple calls still need application-level state management.² The release also changes authorization and adds a framework for extensions.¹

OpenAI released the open-source Codex Security CLI on July 29, 2026.¹ It scans repositories for vulnerabilities, tracks findings across runs, and can check whether fixes work.¹ Teams can also add its security checks to CI/CD pipelines.¹ A TypeScript SDK is available alongside the command-line tool.²

The software used to access the service is open source, but the underlying Codex Security scanner remains in a restricted beta.²³ Teams can prepare an integration, but they still need approval to use the service.⁴³ OpenAI positions it as a way for security and engineering teams to find, confirm, and fix vulnerabilities.⁴ For companies, access to the service is therefore a practical consideration alongside the technical integration.⁴³

Microsoft unveiled Project Perception on July 27, 2026. The security system uses specialized AI agents in three roles: red team agents look for potential attack paths, blue team agents investigate and prioritize risks, and green team agents handle remediation and hardening. Microsoft describes it as a continuously learning defense system. A public preview is planned for early August.¹²

The agents are intended to connect work from finding attack paths through remediation, rather than merely generating alerts.¹ Microsoft also introduced MAI-Cyber-1-Flash, its own cybersecurity-focused model, which is used in the MDASH vulnerability-management system.¹³ It remains unclear how much of Project Perception will be available outside existing Microsoft security workflows. Specific pricing and eligibility have also not been clarified.⁴

OpenAI announced ChatGPT for Academic Researchers on July 29. The free program gives researchers at selected academic institutions access to advanced models and tools through ChatGPT, ChatGPT Work, and Codex. OpenAI plans to start with 10,000 researchers this summer and expand to 100,000 through 2027. The rollout is underway at institutions including the Institute for Advanced Study and École normale supérieure.¹

Access is not open to all researchers: applicants must verify their institutional affiliation and describe their active research and intended scientific use. OpenAI also offers training and hands-on support. The company says data from the research workspace is not used to train models by default. For research teams, the program could lower the cost of using current AI tools, though participation depends on approval.¹²

On July 27, 2026, Moonshot AI released parts of the infrastructure behind Kimi K3. Alongside the model weights and technical report, the company released code for MoonEP and AgentEnv; FlashKDA was already open source.¹ Developed with KVCache.ai, AgentEnv provides isolated environments and supports saving, resuming, and forking agent runs.¹ The released weights are also available through Moonshot’s public repository.²

Moonshot describes AgentEnv as infrastructure for training agents with reinforcement learning.¹ The release gives other teams access to components involved in developing the model, not just its weights.¹² It is not one complete training system: Moonshot identifies AgentEnv, MoonEP, and FlashKDA as separate components.¹ The announcement does not establish how much additional infrastructure a team would need to run them independently.¹

Thinking Machines Lab released Inkling-Small with openly available model weights on July 30, 2026. The model has 276 billion total parameters, with 12 billion active. Its weights are available to download from Hugging Face; users can fine-tune it on Tinker or try it in Tinker Playground. It accepts text, image, and audio inputs and produces text output.¹²

Thinking Machines describes its performance as comparable to the larger Inkling model at about a quarter of the size. That is the company’s assessment, not a claim that the models perform equally on every task.¹ Artificial Analysis scored Inkling-Small at 40 and Inkling at 41 on its Intelligence Index.³ For companies, the question is whether the smaller model’s resource requirements and performance fit their specific applications.¹