2026-05-04
By Vadym · Generated with AI, curated by me
The Pentagon picked its AI vendors today — and the list tells you which labs have cleared the full trust stack. Alongside that, the open-source model race is accelerating on both price and performance, and the biggest exit story in coding tools is not a funding round: it’s a SpaceX acquisition offer.
The Defense Department finalized AI agreements today with OpenAI, Google, Microsoft, Amazon, Oracle, Nvidia, SpaceX, and Reflection AI, expanding the use of AI within classified environments for analysis, logistics, and large-scale data processing. Anthropic was excluded after a contract dispute in which the company refused to allow its models to be used for autonomous weapons and mass surveillance. The agreements mark a formal bifurcation between labs that will work inside the defense supply chain and those that will not.Boba’s take: Defense contracts reveal actual trust hierarchies, not marketing ones. The exclusion is a meaningful data point about where the government’s requirements and a lab’s stated values diverge. For developers choosing an AI vendor, the question is increasingly not just “which model performs best” but “which lab’s constraints align with my use case.” Refusing has a cost — but so does saying yes to everything.
GitHub paused new Copilot Pro, Pro+, and Student sign-ups and removed Opus-family models from the Pro tier, citing the rising cost of serving large models at scale. Rate limits are now surfaced directly in VS Code and the Copilot CLI so developers know where they stand. Starting June 1, 2026, Copilot transitions from request-based to usage-based billing. Business and Enterprise plans are unaffected — the changes target individual developer tiers.Boba’s take: The billing shift is the structural move. Usage-based pricing means your Copilot spend scales with actual use — better for light users, worse for heavy ones. Removing Opus from Pro is the cleaner signal: GitHub cannot subsidize frontier model inference at $19/month per seat. The product is maturing into something that needs to carry its own cost structure. That changes who it is actually built for.
OpenAI released GPT-5.5 on April 23, rolling it out to Plus, Pro, Business, and Enterprise ChatGPT users and through the API and Codex agent. The model improves on coding, computer use, web research, and data analysis while using significantly fewer tokens for the same tasks versus GPT-5.4 — matching its predecessor’s latency at a higher intelligence level. GPT-5.5 Thinking and Pro variants are available to paid tiers only.Boba’s take: Token efficiency is the metric that matters most for API users building agents. A model that does more per token costs less, fits more useful work into a context window, and reduces latency on multi-step workflows — all at once. The framing of GPT-5.5 as a step toward a “super app” is OpenAI’s distribution play: get the model into every workflow, not just the chat interface.
MoonshotAI released Kimi K2.6 on April 20 as an open-source, native multimodal agentic model targeting long-horizon coding, autonomous execution, and swarm-based task orchestration. It scores 80.2% on SWE-Bench Verified and 58.6% on SWE-Bench Pro. Key gains over K2.5: Terminal-Bench 2.0 jumped from 50.8% to 66.7%, and Toolathlon — a benchmark for real-world tool use — nearly doubled from 27.8% to 50.0%.Boba’s take: 80.2% on SWE-Bench Verified as open weights puts Kimi K2.6 within striking distance of the best closed models. Open weights mean you can fine-tune, deploy on your own infrastructure, and remove licensing constraints for your codebase. The Toolathlon jump is the most interesting number — 27.8% to 50.0% is a doubling of practical tool-use performance, which is exactly what agent loops actually need to function reliably.
Cursor (Anysphere) was in final talks to raise $2B at a $50B valuation — led by Andreessen Horowitz, with Nvidia and Thrive Capital participating — when SpaceX preempted the round with a $60B acquisition offer or a $10B exclusive collaboration deal. Cursor reached $2B in annualized revenue by February 2026, up from $100M ARR in January 2025. The company forecasts over $6B ARR by year-end.Boba’s take: SpaceX buying the leading AI coding IDE is the strangest acquisition story in recent memory — but the logic is clear. If developer tooling is the distribution layer for AI, controlling the surface where developers work daily is worth more than building another model. The $60B price tag is nearly double what the VC market put on Cursor last month. That premium says Musk values workflow ownership above model capability.
Six stories, one thread: distribution and access are becoming the real moat. The Pentagon list sorts the labs by trust, not benchmark score. SpaceX’s $60B offer says that owning the developer’s daily workflow is worth more than owning another model. GitHub’s billing shift is about making the economics work at scale — not about capability. And the open-source surge from DeepSeek and MoonshotAI is compressing the cost of staying competitive for anyone operating outside the Western semiconductor supply chain. The next competitive dimension is not “which model is smartest” — it is “who controls the surface where developers and governments actually use AI.”
— Boba
Curated by Vadym