TypeSafe AI Emerges From Stealth With Jev, a Model That Returns Decisions Instead of Text
TypeSafe AI came out of stealth on September 15, 2026, announcing $40 million in seed funding led by DCVC and introducing Jev, the first of what it calls "System One Models." The company was founded by Diogo Almeida — a former OpenAI researcher and co-inventor of RLHF, the technique underlying ChatGPT's original alignment training — alongside Erik Gafni and Sasha Sheng.
Jev's pitch is a deliberate departure from chat-oriented LLMs: instead of generating natural-language text in response to a prompt, it returns typed, probabilistic decisions (booleans, categorical choices, or calibrated scores) in a single forward pass with zero generated tokens. TypeSafe positions this for machine-to-machine use inside automation pipelines — classification, routing, validation, and guardrail checks — where a full conversational LLM is both slower and harder to integrate into deterministic software than a typed decision endpoint. The company claims Jev is up to 193.6x faster and 244.6x cheaper than comparable LLM-based approaches for these tasks, pricing input tokens at $42 per billion, which it describes as roughly 238x lower cost than Claude Fable 5.1.
A follow-up Hugging Face benchmark published this week backs up part of that positioning: in a comparison of 13 answer-verifier systems across 2,018 test items, Jev tied for the top AUC score (~0.736) against a 397-billion-parameter competitor while costing $0.024 per 1,000 calls versus $0.55 for GPT-5.2 — though the same study found Jev's AUC gains didn't fully translate into better end-to-end agent accuracy, a reminder that verifier benchmarks and deployment value aren't the same thing.
Read more — TypeSafe AI
PrismML Launches Bonsai 2 27B, a Ternary-Quantized Model for Edge Devices
PrismML, the Caltech-spinout AI company backed by Khosla Ventures, Cerberus, Samsung, and Google, launched Bonsai 2 27B on September 17, 2026 — its most capable model since emerging from stealth in March 2026 with the first commercially viable 1-bit LLMs.
Bonsai 2 27B is built on a ternary-quantized architecture (weights restricted to three values) derived from Qwen3.8 27B, and PrismML says it retains 98.2% of its full-precision counterpart's performance at a 9x smaller footprint — just 5.9 GB on disk. The company frames the release around "intelligence density" (capability per bit) rather than raw parameter count, and claims 9x less memory use, 8x faster execution, and 5x lower energy consumption relative to standard precision, aimed at making frontier-adjacent capability runnable on smartphones and other edge hardware without a datacenter round-trip. The model adds reasoning, coding, vision, and agentic tool-use capabilities the earlier 1-bit Bonsai 8B lacked.
For developers, ternary quantization at this scale is notable less for the compression ratio itself — 4-bit and lower quantization schemes are already common — than for PrismML's claim of near-lossless quality retention at 1.58 bits per weight (ternary's effective bit width), which if it holds up under independent evaluation would meaningfully shift the cost/quality tradeoff for on-device inference.
Read more — PrismML
GitHub Copilot Adds Cost-Aware Model Tiers, Smarter Code Review, and Sentry Integration
GitHub shipped its September 14, 2026 weekly Copilot release with three developer-facing changes. Auto model selection now offers three tiers — efficiency, balance, and intelligence — letting teams choose how Copilot trades off cost, response quality, and latency when it picks a model on their behalf; all three tiers draw from the same underlying model pool, so the choice only changes selection weighting, not availability. The rollout covers VS Code, Copilot CLI, and the Copilot app.
Copilot code review picked up two workflow improvements: it now resolves comments it previously raised once a follow-up review confirms they were addressed, leaving only outstanding feedback open instead of re-surfacing stale threads, and it suggests commit messages when a developer applies one of its suggestions. Reviews can also invoke shell tools to validate changes before commenting, and "Lite" reviews now merge findings from multiple review agents into a single pass.
Separately, GitHub added a Sentry integration to the Copilot app, letting developers jump directly from a Sentry crash report into a Copilot session with the error, stack trace, and surrounding context already loaded — shortening the path from "alert fires" to "PR opened" for production incidents. GitHub also confirmed the deprecation of several older Copilot models across chat, inline edits, agent mode, and completions ahead of an October 19, 2026 cutoff, with migration guidance available for Business and Enterprise customers.
Read more — GitHub Changelog
OpenAI Ships Codex CLI 0.155 With Experimental Voice Mode and Touch ID for MCP
OpenAI's Codex CLI moved through two releases this week. Version 0.155.0 (September 17) is the larger of the two, adding experimental /voice conversations with live transcripts directly in the CLI, Touch ID verification for MCP requests on macOS (gating tool calls from MCP servers behind a biometric prompt rather than a blanket approval), configurable daemon update schedules, and support for pulling AWS Bedrock credentials from a configured command rather than static environment variables. A same-day patch, 0.155.1, fixed a regression where reasoning summaries were disabled by default in a way that caused request rejections from providers that don't support the summary field.
The prior release, 0.154.0 (September 10), had already added availability for GPT-6-Astra inside Codex, experimental Git worktree support for isolating agent changes per branch, inline question answering during a session, a native Windows daemon, and Vim replace-mode support in the terminal UI. The accompanying Python SDK release (0.154.0) added max and ultra reasoning-effort levels and an ExternalMessage type for injecting messages into a running session's history.
Separately, OpenAI confirmed GPT-5.5 will retire from ChatGPT, Work, and Codex across all plans on October 14, 2026, with users redirected to GPT-5.6-Sol; the GPT-5.3-Codex-Spark research preview was also discontinued across desktop, CLI, and IDE surfaces on September 14.
Read more — OpenAI/ChatGPT Changelog
Hermes Agent Closes Out a Session-Storage Reliability Campaign
NousResearch's Hermes Agent shipped three patch releases in the two weeks after its "Pantheon" v0.21.0 release, all focused on session-storage reliability rather than new features. v0.21.1 (September 7) rolled up roughly 632 merged PRs of general codebase modularization and performance work. v0.21.2 (September 11) targeted database corruption issues introduced in v0.21.0, closing 44 issues across six PRs and hardening multi-profile isolation, plus eliminating a class of desktop-backend process spawn storms. v0.21.3 (September 14) followed with a further ~338-PR rollup, fixing remote dashboard sessions expiring on rapid refresh and closing writer-handle leaks in state.db for long-lived agent processes.
Taken together, the pattern is a lab working through the operational debt of a large feature release (Pantheon added bundled Bot Mode, persistent scheduled-job memory, live subagent orchestration, and an MCP management dashboard) rather than shipping new capability — worth noting for teams running Hermes Agent in production who hit session-storage instability on v0.21.0 and should now update to v0.21.3 or later.
Read more — GitHub
Links & Sources
- Introducing System One Models & Jev
- TypeSafe AI Emerges From Stealth With $40M in Funding
- PrismML Launches Bonsai 2 27B, Its Most Capable Model Yet
- PrismML hopes its tiny LLM will change how we all use AI
- Inside the JEV Ecosystem: 13 Answer Verifiers on One Test Set
- GitHub Copilot weekly releases — September 14
- OpenAI/ChatGPT Changelog
- Hermes Agent Releases
You May Also Enjoy
