Three model moves this week all pulled the same lever: a capable coding-and-agent model got cheaper and moved closer to where you already work. xAI shipped Grok 4.7 and GitHub put it in every paid Copilot tier the same day at $2/$6 per million tokens — a frontier-adjacent coder now one dropdown away in the IDE you already use. StepFun opened Step 5 Preview, a 600B mixture-of-experts model that scores like Kimi K3 Max for roughly a seventh of a US frontier model's price, with open weights due Oct 15. And at its Apsara conference, Alibaba unveiled a new AI chip and a road toward 5-to-10-trillion-parameter Qwen models. If you run an agent or ship code with one, the cheapest capable model you picked in the summer is probably not the cheapest capable model today.

Here's the whole edition in one screen — the three moves, and the one thing to do about each:

The through-line is a single, founder-friendly direction: the price of a capable coding-and-agent model keeps falling, and the distance from that model to your editor keeps shrinking. The discipline that pays off is the same one we keep coming back to — keep every layer swappable — so that acting on any of this is a config change, not a migration.

1. Grok 4.7: a cheap frontier-adjacent coder, one dropdown away#

The move most likely to change your afternoon is the most convenient one. On Sept 21, 2026, xAI released Grok 4.7 — its most capable coding and agentic model — and GitHub made it available the same day across every paid Copilot tier: Pro, Pro+, Max, Business and Enterprise. You pick it from the model selector in VS Code, Visual Studio, the Copilot CLI, the Copilot cloud agent, JetBrains, Xcode and Eclipse — and it's on the xAI API directly for anything you build outside the editor.

Pricing is the headline: $2 per 1M input tokens and $6 per 1M output, unchanged from Grok 4.6 and well under the top US frontier models. Reported benchmark gains put it around 38% on Terminal-Bench and 46% on CursorBench — but those come from a single aggregator, so read them as claims, not settled fact. It still trails the frontier on the hardest reasoning; the pitch is cheap and good enough for agentic coding, wired into the tool you already open.

What it means. The convenience is the point: there's no new SDK, no new account — just a different entry in a picker you already use. But cheap-per-token is not the same as cheap-per-task. As VentureBeat noted, Grok 4.7 tends to consume more tokens per job, which can quietly eat the per-token savings. So don't switch on the price sheet — switch on a measured comparison. Run one real task through your current model and through Grok 4.7 and compare end-to-end cost and output, the same discipline we laid out for choosing an open-weight coder to self-host. If you're standing up Copilot or Claude Code fresh, our IDE setup guide still applies — the model picker is where this lands.

2. StepFun Step 5 Preview: frontier-tier scores at one-seventh the price#

The story with the biggest number attached to it is the cheapest. On Sept 20, the Chinese lab StepFun opened Step 5 Preview — a 600-billion-parameter sparse MoE with ~27B parameters active per token, a 1M-token context and multimodal input, built for long-horizon agentic, coding and analysis work. API access opened the same day at $1 per 1M input and $2.70 per 1M output, with a 95% cache discount ($0.05 cached input).

The reason to care is the price-for-quality ratio. Artificial Analysis scores it 44 on its Intelligence Index — the same as Kimi K3 Max — at roughly one-seventh the price of a US frontier model at a comparable tier. And StepFun has put a date on self-hosting: full open weights on Oct 15, 2026.

What it means. For high-volume, latency-tolerant or long-context work — batch analysis, RAG over big corpora, agent loops that chew through context — a model that scores like Kimi K3 Max at a seventh of the cost is a serious line item on your inference bill. The sparse activation matters too: at ~27B active, its serving cost behaves closer to a mid-size dense model than its 600B total suggests, which is exactly the economics we traced in the September LLM API pricing breakdown. Price-test the hosted API now to see whether the quality fits your workload, and put Oct 15 on the calendar to re-run the self-host math — the same head-to-head we did for DeepSeek vs GLM vs Qwen self-hosted coders will want a new column.

3. Alibaba's Apsara roadmap: the supply signal under the cheap models#

The macro story you can't act on but shouldn't ignore came out of Hangzhou. At the Apsara Conference, Alibaba CEO Eddie Wu unveiled the Zhenwu V900, calling it the most powerful AI chip in China — about 3x the performance of May's Zhenwu M890, able to scale to 500,000 cards per cluster, with mass production planned for Q1 2027. Wu also said Alibaba is training toward ASI: Qwen 4.5 and Qwen 5 are targeting 5-to-10 trillion parameters, up to four times today's ~2.4T Qwen 3.8 Max, and the company plans to push global data-center capacity past 20GW by 2032.

What it means. None of this is a purchase decision this quarter — it's a supply signal, and it points one way. The cheap, capable, open-weight models your budget tier increasingly runs on — Qwen, Kimi, GLM, and now StepFun — come out of exactly this kind of chip-and-model buildout. A domestic chip at 3x last quarter's performance and a model roadmap 4x today's flagship means that pipeline is accelerating, not slowing. The practical takeaway is portability: keep an open-weight Qwen path in your model plan, treat every model choice as swappable, and — as we've said through a summer of falling prices — don't lock multi-year compute commitments at today's rates on the assumption the floor has been found. It keeps dropping.

The one-week picture#

Three moves, one direction: a capable model got cheaper (StepFun, ~7x under the frontier), got closer (Grok 4.7, in every Copilot tier the same day), and got a supply chain behind it that keeps both trends going (Alibaba's chip and 10T-parameter road). For a team of one the response is a single motion — re-run your model bake-off — because the default you picked in the summer has almost certainly been undercut, and the only way to know by how much is to measure it on your own task. Keep every layer swappable and this stays a config change. If getting found is the other half of your plan, the playbook for getting cited by AI answer engines is still the cheapest moat a solo founder can build — but first, ship on the cheapest capable model, and this week that ranking changed.