Every Major Model & Agent-Harness Release, August–September 2026

Models

Prices are standard (non-batch) API rates. Our cost calculator carries most of these, so you can model them against your own cache-hit rate, which is where most of the real difference between them shows up.

Jev — the one that isn't a chatbot at all

Released in early access on 2026-09-15, Jev is a different model class, not a cheaper LLM. You send a state plus typed questions — Choice (pick an option), Score (rate against ordered levels), Noul (probability a yes/no statement is true) — and it returns every answer in one parallel pass with a probability attached, in 70–500ms. It cannot return a value outside your schema and cannot produce a type error. It also cannot write a sentence, explain itself, or take an image.

Pricing is the headline for agent builders: $0.042 per million input tokens and free output, against $10/$50 for the three frontier flagships above. That makes per-turn classification — routing, tool-call gating, triage — effectively free, and it is why this belongs in a model roundup even though it competes with none of the others directly. The catch is that "cannot hallucinate" is a claim about types, not truth: a question fitting none of your options still comes back with one, observed at 0.31 confidence. Full detail, setup and measurement discipline: our Jev section.

Gemini 3.8 Flash — the workhorse benchmark results kept pointing to

Released 2026-09-02 at $0.75 / $3.75 per MTok with a 1M context. This model doesn't have a video page of its own yet, but it keeps showing up in other people's tests. In IndyDevDan's five-benchmark ranking it placed second on Deep SWE (long-horizon software engineering), behind only GPT-6 Astra, and was a strong runner-up on the Omniscience hallucination benchmark. The trade-off is time. It takes more steps and more tokens to get there. What to do: it's the obvious candidate for the "workhorse" slot in a model stack. Route long, low-stakes tasks to it, and keep a frontier model for work where a wrong answer is expensive.

Grok 4.7 and 4.6 — same price, and the same cliff at 200K

Grok 4.7 shipped on 2026-09-21 at exactly Grok 4.6's price, with a larger base model, a longer reinforcement-learning run and a knowledge cutoff of May 2026. xAI's own numbers put it ahead of 4.6 on every benchmark it published — Terminal-Bench 4.0 37.6% against 20.3%, DeepSWE v1.1 71.0% against 65.2% — and say it was trained to natively understand the Grok Bot harness. Those are the vendor's results; we have not seen independent agent benchmarks yet. A fast variant runs at twice the output speed for twice the price.

Grok 4.6 was xAI's flagship from 2026-08-12, with a 500K context — as is 4.7. The pricing detail to plan around: below 200K tokens a prompt costs $2 / $6 (cached input $0.50), but once a prompt reaches 200K, the whole request bills at $4 / $12, not just the tokens over the line. An agent that lets its context drift past 200K doubles its cost on every turn after that. On Automation Bench, which penalises guardrail violations, Grok 4.6 ranked near the top. xAI's hosted agent product, Grok Bot, runs on xAI models, but its docs have never named which version — see our Grok Bot section.

Muse Spark 1.3 — Meta's agentic API model, with a cheap tier that trains on your data

Released 2026-09-02 on Meta's API (and in Muse Code) as muse-spark-1.3: $1.25 input / $0.15 cached / $4.25 output per MTok, 1M context. Meta says it uses about 20% fewer tool calls and 25% fewer tokens than 1.2 on its internal comparisons, and is tuned for longer multi-step work. There is also a muse-spark-1.3-contributor variant at $0.10 / $0.002 / $0.20, which Meta labels "used to improve our products". What to do: treat the contributor tier as a public channel. It's fine for open-source code or synthetic tests, but never for client data, credentials, or anything under NDA.

DeepSeek V4.1 Flash — the old model names now route to it

Released 2026-09-10 as deepseek-flash: a 552B-parameter MoE with native vision and a 1M context. Peak: $0.30 / $1.20. Off-peak: exactly half. Cache hits: $0.006 peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday. The migration detail: V4 Flash and V4 Flash Vision Exp are retired. The names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but they route to and bill as V4.1 Flash. If you pinned the old name for reproducibility, you're already on a different model. Measured on real builds, it cost 4–6× less than Astra but took 3–5× longer.

MiMo-V2.6 — Xiaomi's cheap, MIT-licensed pair

Xiaomi released MiMo-V2.6 Pro and Flash on 2026-09-21, with MIT-licensed weights on Hugging Face (XiaomiMiMo/MiMo-V2.6-Pro-RL, -Flash-RL, plus a distilled Qwen 9B). On OpenRouter, Pro lists at $0.435 / $0.87 and Flash at $0.14 / $0.28, both with a 1M context and near-free cache reads ($0.004 and $0.003 per MTok). That combination — permissive licence, frontier-length context, very cheap repeated prompts — is exactly what long-running agents want, but the model is days old and independent agent benchmarks are thin. Test it on your own workload before routing volume to it.

GLM-5.3-Flash and FlashX — a fast open-weights option

glm-5.3-flash (launched 2026-08-26, MIT licence, weights on Hugging Face) is the first natively multimodal GLM-5 model: 320B total parameters, 18B active, 1M context, 128K max output. glm-5.3-flashx followed on 2026-09-18 at up to 200 tokens/second. Z.ai's docs say FlashX isn't on the GLM Coding Plan yet, while Flash is. In the benchmark roundups it's a strong budget option for coding, but weak on hallucination, so pair it with verification, not trust. Z.ai's docs page doesn't list per-token prices; OpenRouter lists Flash at $0.15 / $0.50 and FlashX at $0.37 / $1.25 (checked 2026-09-24). The full GLM-5.3 (2026-08-18, custom licence) is $0.84 / $2.64 there.

Harnesses: what changed in how agents load and run