Issue #5 · July 11, 2026 · ~5 min read

Four coding models in one week

Welcome back to Zero to MVP Weekly. Same deal: one signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. Let's go.

1. The signal of the week

OpenAI, SpaceXAI, Meta, and Tencent all shipped coding models this week. The story isn't any single launch — it's that list price quietly stopped being a number you can compare.

On July 9, OpenAI made GPT-5.6 public in three tiers: Sol, the flagship, at $5 per million input tokens and $30 per million output; Terra, the everyday default, at $2.50/$15; and Luna, the cheap fast one, at $1/$6 — all with a 1M-token context window and 128K max output. Note how it arrived: after roughly two weeks of preview restricted to government-approved organizations. One issue after Fable 5's export-control blackout, a government-coordinated launch is starting to look like standard procedure, not an exception. The same day, all three landed in GitHub Copilot's model picker — across VS Code, the CLI, the cloud agent, JetBrains, and Xcode.

The rest of the week filled in the market around them. On July 8, SpaceXAI shipped Grok 4.5 — its first model co-built with Cursor since the acquisition — at $2/$6, undercutting everyone in its class; its whole pitch is token frugality, claiming roughly 60% fewer output tokens than Opus 4.8 on the same tasks, and independent rankings put it around fourth on agentic work but far cheaper per task than the frontier. On July 9, Meta entered the race with Muse Spark 1.1 and a new Meta Model API aimed squarely at AI coding tools. And on July 6, Tencent released Hy3 — a 295B-parameter MoE that runs like a 21B, with 256K context — under Apache 2.0, free on OpenRouter through July 21.

Here's the builder takeaway, in two parts. First, the shape of launches changed: tiered families and model pickers mean routing between models is now the default workflow, not an optimization — the "route, don't switch" idea this newsletter has pushed since #002 just became the vendors' own UI. Second, the math changed: models now differ wildly in how many tokens they burn to finish the same job, so comparing $/M tokens across models is comparing prices in different currencies. The only number that matters is cost per completed, accepted task — and almost nobody measures it.

One caveat to keep us honest: every benchmark in this story is vendor-picked. OpenAI leads on the agentic evals it highlights, while Fable 5 still leads SWE-Bench Pro by a wide margin — and OpenAI's own audit found about 30% of SWE-Bench Pro tasks problematic. Treat the numbers as marketing until they survive your own tickets.

2. My take

The only real price of a model is what it costs to finish your task. Everything else is marketing.

Four labs published their price lists this week, and the numbers look comparable — dollars per million tokens, neat columns. They aren't comparable. Models differ several-fold in how many tokens they burn to finish the same job: Grok 4.5's entire pitch is that it says less; early GPT-5.6 users are complaining about the opposite. A cheap model that needs three retries and your intervention is an expensive model.

I saw this first-hand this week running Grok 4.5 in Cursor. On small, clearly defined tasks it finished for cents and I had nothing to complain about. On anything that needed architectural judgment, it burned tokens and my time — the two costs no price list shows.

So my advice is the light version, not the lab. Pick two or three of your real, recurring tasks. Run each through two models — your current default and one cheap challenger. Write down three numbers: what the session cost, how many times you corrected it, whether the tests passed. A week of that gives you a routing rule you'll actually trust — their benchmarks measure their tasks; your bill comes from yours.

And the skip: don't build a full eval rig before you have revenue. That's a product (one I'd pay for — see the MVP idea below), not a habit. A note in a text file beats an unbuilt dashboard.

3. Tool worth trying

Grok 4.5 inside Cursor.

I tested Cursor's first native SpaceXAI model myself. It is built for coding and agent loops, costs $2/$6 per million tokens — or $4/$18 for the fast variant — and is available in Cursor desktop, web, and CLI across all plans.

On smaller, clearly defined coding tasks, it works well. I was able to run it successfully, and for this kind of work it can be a practical and relatively inexpensive option. But once the task requires deeper analysis, architectural decisions, or multi-step planning, it still falls behind Opus, Fable, and GPT-5.5/5.6.

A useful test is to give it the same real refactoring task as your current default model, including tests, and compare not only the final code but also the plan, the number of corrections required, and the reported token usage. My impression is that it is already useful for focused implementation work, but the hardest tasks still benefit from a frontier model. It is also very new software, so I would be cautious with critical projects and apply the usual privacy review before using it with client code.

4. MVP idea of the week

A cost-per-task meter for coding agents.

Problem: four new models in one week, tiered pricing everywhere, and token appetites that differ several-fold per task — nobody can honestly answer "which model is cheapest for my work," because list price stopped correlating with real cost.
Solution: a harness that replays a team's own recurring tasks across N models via BYOK and reports cost per accepted result: tokens, retries, human interventions, test pass rate, diff size. Output: a routing policy file the team's model picker can actually use.
For whom: small teams with real agent bills; agencies pricing AI work; reviewers who want numbers instead of vibes.
Minimal v1: a CLI wrapping one agent runtime (Copilot CLI or Claude Code), five task templates, CSV out, a markdown report. No dashboard.
Demand check: run your own ten real tickets through four models and publish "the same work, four bills" with real numbers; count who asks to run it on their tasks.
Why now: GPT-5.6's tiers, Grok's token-frugality pitch, and Copilot's picker all landed within the same seven days. Routing became the interface before anyone shipped the meter.

5. Useful links

Simon Willison — The new GPT-5.6 family: Luna, Terra, Sol — the honest first read: pricing, contested benchmarks, and why token counts now matter more than token prices.
TechCrunch — SpaceXAI releases Grok 4.5 — the price war's sharpest move, and the token-efficiency argument in full.
GitHub Changelog — Copilot app available to all — a maintained agent runtime, free tier plus BYOK; the fastest way to act on this issue.
VentureBeat — Tencent's Apache-licensed Hy3 — the open-weight side of the same price war: 295B that runs like a 21B, free to try this week.
GitHub Changelog — npm v12 install-time security — off-signal but do it anyway: install scripts no longer auto-run; migrate deliberately before your build breaks.