Need to know
- Anthropic filed confidential IPO papers with the SEC, flipping prediction markets to give it a 76% chance of listing before OpenAI.
- OpenAI models including GPT-5.5, GPT-5.4, and Codex are now generally available on Amazon Bedrock at direct API pricing.
- Microsoft Project Polaris, an in-house MoE coding model, will replace GPT-4 Turbo as GitHub Copilot's default engine by August 2026.
- Palo Alto Networks closed its acquisition of AI Gateway startup Portkey, adding a control plane for governing autonomous agents.
- JetBrains open-sourced Mellum2, a 12B MoE model purpose-built for software engineering tasks in multi-model pipelines.
New Releases
Amazon has made GPT-5.5, GPT-5.4, and Codex generally available on Bedrock, letting AWS customers run OpenAI's latest models inside existing AWS environments.
- Pricing matches OpenAI direct rates, meaning AWS adds no markup on tokens; Codex runs pay-per-token alongside the other models in the Bedrock catalog.
- The launch follows an expanded AWS-OpenAI partnership, giving enterprises a path to OpenAI models without leaving AWS IAM, logging, or compliance tooling.
Microsoft announced Project Polaris at Build 2026, an in-house mixture-of-experts coding model that will displace GPT-4 Turbo as GitHub Copilot's default by August 2026.
- Polaris runs on Microsoft's Maia AI accelerators inside Azure, reducing per-token cost and giving Microsoft margin control independent of OpenAI pricing.
- The model uses language- and framework-specific sub-modules, a structural bet that specialized routing beats a single large general model for code tasks.
JetBrains released Mellum2 under Apache 2.0, a 12B MoE model covering code generation, debugging, tool use, and agentic coding designed to serve as a fast focal component in multi-model pipelines.
- Mellum2 replaces the original 4B dense Mellum, expanding from completion-only to general software engineering including multi-step reasoning and function calling.
- The 'focal model' framing positions Mellum2 as the fast, specialized node in an orchestrated stack rather than a standalone frontier competitor.
Anthropic gave approximately 150 organizations worldwide access to Claude Mythos as part of an expanded Project Glasswing cybersecurity initiative, while committing to eventual public release.
- Mythos was first shared only with Apple and a handful of select partners in April; today's expansion adds organizations across more countries still under controlled access terms.
- The model's primary use case remains cybersecurity, specifically rapid identification of weaknesses in computer systems, which Anthropic says has prompted global concern about dual-use risk.
Funding
Anthropic filed confidential IPO documents with the SEC on June 2, immediately shifting Kalshi and Polymarket prediction markets from giving OpenAI a 76% chance of listing first to giving Anthropic a 76% chance, a signal that institutional investors see Anthropic's $47B revenue run rate and $965B valuation as the cleaner public story.
Palo Alto Networks completed its acquisition of Portkey, an AI Gateway startup, establishing a central control plane inside Prisma AIRS to monitor, orchestrate, and govern autonomous agents, signaling that the security industry now treats AI traffic routing as a mission-critical infrastructure layer rather than an app-level feature.
NP Company closed a €6M pre-seed backed by Mistral AI's founders to build AI for engineering workflows, a small round that matters because it shows Mistral's leadership actively seeding vertical AI applications in Europe's industrial sector.
Case Studies
Cato Networks reported cutting time-to-protect for newly discovered CVEs to 45 minutes using agentic threat intelligence, compared to the industry norm of days to weeks tied to appliance patch cycles.
- Traditional appliance-based security depends on vendor patch cycles that leave organizations exposed for days after disclosure; Cato's cloud-native SASE architecture lets AI agents push protection without waiting for hardware updates.
- The 45-minute figure represents a claimed world record for vulnerability mitigation speed, a benchmark that will pressure legacy perimeter vendors whose patch-and-deploy model cannot match agentic response times.
Trending on X
- Anthropic IPO filing flips prediction markets Polymarket and Kalshi both swung from 76% OpenAI to 76% Anthropic within hours of the SEC filing, and the debate centers on whether Anthropic's higher reported revenue makes it the cleaner narrative for public investors despite OpenAI's brand dominance.
- AI coding agents vs. human release bottlenecks Ethan Mollick flagged a new paper showing autocomplete tools (Copilot) produce 2.2x more code, local agents 7.4x, and remote coding agents 17.3x, but actual software releases only rose 30% because human review and QA are now the binding constraint.
- GitHub Copilot token billing developer backlash The switch from flat-rate to AI Credits billing on June 1 is generating sustained anger on X as developers discover their effective monthly cost has risen beyond the unchanged base subscription price, with power users hitting overages on day one.
- Microsoft building its own model to replace OpenAI Project Polaris is drawing significant discussion because it reframes the Microsoft-OpenAI relationship: the $13B investor is now training a competing model to run inside its own flagship product, raising questions about OpenAI's long-term revenue concentration.
- Grok destroyed simulated world in four days Emergence AI's experiment placing leading AI models in charge of a simulated society is circulating widely, with Grok reaching total societal collapse in four days versus other models surviving longer, reigniting debate about safety alignment in agentic systems.