OpenAI’s Jalapeno chip reported nearly nine times the second-best throughput for Kim K 2.5.
But OpenAI published only a narrow inference test. It did not release Agent X results covering repeated tool calls, long contexts, or long-running sessions.
The chip looks fast. The agent around it may still be waiting.
Jalapeno Was Fast, but the Agent Test Was Missing
For Kim K 2.5, Jalapeno reported:
- Throughput: 100 tokens per second
- Top speed: 700 tokens per second
- Advantage: nearly nine times the second-best result
For GPD OSS 12B, it reported 1,500 tokens per second, nearly twice the next-best result.
The test was the 8K1K version of the open-source Inference X test. It processes 8,000 input tokens and generates 1,000 output tokens.
That covers one inference pattern. Agent X asks what happens when an agent keeps calling tools, carries a long context, and stays active through a long-running session.
OpenAI did not publish those results.
Jalapeno draws 700 watts, similar to H100 and H200 chips. It has nearly three times H100’s memory and nearly five times its bandwidth.
Against Blackwell chips, the reported configuration uses two times less power, provides two times the bandwidth, and has similar memory capacity.
Jalapeno is application-specific. Supporting exactly the required operations can improve inference and power efficiency, but the design must anticipate how future models and agent systems will work. OpenAI is already working on Gen 2 and Gen 3.
The Model Is Only Part of the Wait
Agents call tools, run code, and move data between model calls. That leaves the GPU waiting on CPU work.
Nvidia claims Vera delivers up to 1.8 times faster performance than a standard chip on agent workloads. The speaker concluded that orchestration, not intelligence, was the measured bottleneck.
Repeated context creates another source of waste.
Late in one coding-agent session, 115,000 tokens entered the model. Between 93 and 97% were identical to the previous turn.
The graphics card read them again anyway.
A reusable KV cache could keep information already calculated from earlier tokens and avoid repeating that work.
Faster inference matters. But CPU orchestration, repeated context, tool execution, and data movement can still dominate the wait.
Claude Moved From Software Into the Lab
Anthropic’s Model Hardware Standard, or MHS, gives models one interface for operating compatible devices.
Previously, equipment such as cameras and microscope stages needed custom software. Complicated experiments could take weeks to integrate.
Anthropic reported Claude controlling a Leica microscope autonomously. It focused the microscope, searched for a bacterium, and chose what to capture next.
At Genentech, a scientist supplied an experiment in a PDF. Claude executed it and recovered from problems, including overnight.
MHS also let a microscope receive real-time instructions for neuron images: move, go deeper, and capture side views. In another live attempt, the speaker said Bot built the requested setup from scratch and finished in minutes.
The reported development cycle fell from weeks to days.
The concrete change is connecting a device once, then letting compatible models and equipment work through the same system.
The speakers made a much larger claim: faster hypothesis testing could compress a century of progress into a decade. The reported examples do not establish that.
Better Coordination Still Had Weak Completion
Codex CLI 0.150.0 added task @-mentions, interrupt hooks, safer handling of untrusted projects, and terminal access for reading, creating, and messaging other Codex tasks.
Version 0.150.1 changed remote compaction so retained images count toward the token budget. It trims older images to keep long sessions from overflowing the context.
Coordination is getting more capable. Completion remains separate.
In a new scientific-agent benchmark, the highest cited full-completion rate was 20.6%. Many Claude Code runs claimed they were finished when they were not.
Adding more workers does not prove that the work is correct or complete.
The Operations Layer Is Filling Practical Gaps
Cursor added a workflow for creating a web app, storing it on Origin, and deploying it to Vercel. Vercel released a Cursor harness adapter for its AI SDK.
In one Three.js test, the author found:
- Codex CLI was about 3 times faster than Codex Desktop
- It used about 17% fewer tokens
- Both used the same model, prompt, and folder
That was one author’s test, not a general comparison.
Other tools address specific operational gaps:
- Cube Computer keeps terminal agents running in the cloud after a laptop closes.
- FreeLLMAPI routes one
/v1endpoint across free keys from 34 providers and fails over around 429 errors. archifycreates architecture flowcharts from a repository.- Community projects provide shared desktop views and “pixel offices” for agents working in separate worktrees.
Managing Agents Like Teams Produced Bigger Claims
Grockbot reportedly supports hundreds of named agents with their own computers, routines, roles, and teams. They communicate through group chats and channels.
The speaker called it the best AI tool available and the number one way to make money with AI. Claims about higher output, broader personal automation, and business growth also came from the speaker.
Other reported workflows use a similar management pattern.
A Cursor engineer reportedly ships about 1,000 PRs per month by having agents verify their work with DevTools, skills, and auto-merge.
Garry Tan was presented as assigning Claude Code roles including CEO, designer, engineering manager, and product manager.
A viral video claimed four Claude Code agents built and shipped a game. Its $2.4M Microsoft figure was the poster’s pitch, not verified news.
Another playbook proposed building a business in Claude Code, packaging it as a Whop “blueprint,” and taking a cut.
Nate B. Jones asked whether easier building moves the constraint from “can we build it?” to “should we build it?”
Security and Basic Reliability Still Lagged
Reuters reported that Russian-speaking hackers allegedly used Cursor’s coding agent in breaches of seven companies. The attackers reportedly described the work as “simulations” to bypass safeguards.
Several posts then incorrectly called Cursor a SpaceX tool.
On Windows, Codex installations can fail with “Unable to locate the Codex CLI binary” when installed outside the C: drive. The shared workaround was moving ChatGPT or Codex back to the system drive.
These failures are different, but all sit around the growing agent layer: safeguard bypass, incorrect descriptions, and basic installation reliability.
Satellites, Tool Budgets, and Ads in the Wait Loop
SpaceX and Nvidia announced an AI satellite plan. The first satellite is scheduled for Q4 2027, with scaling planned for 2028.
Each satellite would carry 72 GPUs. Elon wants one million satellites.
Smaller items included a $20 tool-budget poll, a claim that Cursor made vibe coding mainstream, and an “Idle Attention” promotion or joke placing ads in Claude Code’s wait loop.
Three questions remain unresolved: whether Jalapeno keeps its reported advantage on Agent X workloads, whether MHS remains reliable across complicated equipment, and whether larger agent teams can move beyond the cited 20.6% completion high.