[AINews] TypeSafe/Jev at >$100M ARR, $7.5B valuation 3 weeks after launch

What it claims

This is a roundup newsletter, so its value is in the specific figures it attaches to each item. The lead claim is that "decision models" became a product category. These return typed answers (a probability, a pick from a list, a score against levels) in one forward pass, and Jev is the reference everyone benchmarks against. TypeSafe's ARR figure comes from a Sequoia "leak" of >$100M in the first week, not from audited disclosure. The authors mention accusations of astroturfing and say their own Jev podcast was authentic, which means Latent Space has a prior coverage relationship with the product.

Evidence

Decision models

• Perplexity pplx-decider-v1.1-27b: 94.5% on Decision Bench across 1,071 cases, at $0.017 per 1K decisions. This is the vendor's own claim of top accuracy.
• Cloudflare clef: clef-flash is cheaper than Jev, and clef is about 2x faster overall. clef-omni accepts audio, video, image and text.
• Microsoft-Decision-1: An early evaluator reports that decision models still struggle with consistency and complex decisions.
• Train your own on Qwen3.5-0.8B: accuracy rose from 37% to 65% in 60 steps, about 10 minutes on 4GB.
• LangChain routing: sending each task to the cheapest adequate model cut median Open SWE cost per task by 64%.
• Apple/CMU SSR (Selection-based Structured Reasoning):
• Method: six natural-language strategies are scored by length-normalized log-likelihood in one batched pass with a shared KV cache.
• Latency: per-turn reasoning latency falls >90%, but end-to-end latency per question falls only 28–54%.
• Accuracy: on Qwen3-VL-4B with GRPO, average success is 61.37% vs 61.25% for TAPO+GSPO, which is parity rather than a gain.

Agent teams (Vals AI, Vibe Code Bench)

• Teams cost 1.8–5.1x more than single agents.
• Only GPT-6 Sol at medium effort improved significantly, by +7.3 points.
• Opus 5.5 at max effort averaged about 6.8 subagents and about 1,140 subagent tool calls per app, with no significant gain. Opus ran sequential waves, while Sol delegated in parallel along architectural lines.

Prime Agent Rust rewrite

• Built by a swarm of 2,000+ agents using 10K+ sandboxes and 200B+ GLM-5.3 tokens over two weeks.
• Reaches usable input about 13x faster and uses 83% less startup memory.

Models

• StepFun Step 5 Preview: 600B total / 27B active MoE with 1M context. Scores 33.89 on the Hermes Index, equal to GPT-6 Luna.
• Solar Mini 4: 35B MoE with 3B active, 524K context, 208 tok/s. AAII score of 24, the best at 3B active.
• Gemini 4 Argon: 77.9% on DeepSWE v1.1 vs Opus 5.5's 74.2%.
• Claude Haiku 5.5: #30 on WebDev at $0.10/$0.50, the same price as GPT-6 Luna and 6 points higher.
• ARC-AGI-3: a new high score of 59.17%.
• Multi-turn editing (30 chained edits):
• Ideogram 4.5 and FLUX 3 leave 95%+ of the image untouched on small edits.
• GPT Image 2.5 Sunburst keeps only about 20% unchanged per turn and drifts.
• Nano Banana 2.1 gradually darkens the image.

Systems

• TRL v1.15: the fused LM head is now on by default.
• On Gemma 3 1B, GRPO max sequence length rises from 28K to 114K and DPO from 10K to 59K.
• Peak memory at 8K falls 52–82%, and training is up to about 11% faster.
• vLLM on Vera Rubin: more than 7.8x GB200 throughput on MiniMax M3 at matched interactivity.
• SGLang: up to 20% faster FP8 MLA at 128K context, plus 5.9% end-to-end from MoE tail fusion.

How to use it

OpenAI Decisions API

• Three request types: probability that a condition is true, pick from a list, or score against levels.
• Accepts text and images and runs on GPT-6 Luna.
• Costs $0.10/M input tokens with no output charge.

Other decision-model options

• clef weights are on Hugging Face.
• Liquid d1 is on Vercel AI Gateway, with vision support for classify, route and score tasks.
• vLLM Semantic Router Decision 2.0 answers several questions about one input in one pass and returns per-option probabilities.
• Unsloth has a free notebook that turns Qwen3.5-4B into a decision model on 8GB of VRAM.

Claude Managed Agents dynamic workflows (public beta)

• Enable with multiagent_20261001.
• Up to 1,000 agents per run.

Other availability

• Opus 5.5 fast mode bills against usage credits and is not included in subscriptions.
• Qwen-Image-2.1-Turbo (7B, 8-step 2K generation) loads via the Diffusers QwenImage21Pipeline.
• Step 5 open weights are due October 15, and it is free on Nous Portal for one week.
• Solar Mini 4 is free in Cline.
• Gemini 4 Argon goes first to 650+ Fairwind Program defenders, at $2/$10 per M tokens.

Limits

• Nearly every number is a vendor or single-tweet claim, linked by handle and not independently verified. The page explicitly flags these as early or unconfirmed:
• "up to 10x faster" for the Decisions API (OpenAI's own figure)
• the Vera Rubin numbers ("early results")
• the SemiAnalysis 3.2x profit-per-GW figure
• the "Carbon" checkpoint report from Business Insider
• Step 5's max output was corrected to 64K tokens.
• Step 5's #1 on OpenRouter Trending measures usage, not quality.
• Anthropic warns that dynamic workflows can use a lot of tokens and advises starting with scoped tasks. Combined with Vals' 1.8–5.1x team cost and mostly null gains, this argues for measuring before fanning out.
• Codex had reported daylong outages, and Composer predictions are Pro-only.
• DeepSeek retrieval has periodic weak spots tied to where a token lands relative to the compression stride. V4.1's stride of 2 reduces the effect but does not eliminate it.