Today on the Brief: The Agent Price War

DeepSeek pushed V4 Pro 0813 to general availability this week — no blog post, no press release, no changelog entry. Just... it was there. That's not how you'd expect a flagship model launch to go, and that's kind of the point with DeepSeek at this stage — they let the model speak instead of the marketing.

And the model has something to say. Its agentic coding benchmark — DeepSWE — jumped from 12.8 to 62.7 in one release. That's not incremental, that's the model getting fundamentally better at multi-step tool use and code execution. Pricing is still absurd — around forty-four cents per million input tokens, eighty-seven cents output — and DeepSeek has already warned a price increase is coming, they just haven't said when.

The same lab that shipped Grok Bot yesterday didn't slow down — xAI, under SpaceXAI now, released Grok 4.6, built specifically for long-running agents: research, coding across a whole codebase, or turning an idea into a working app without you babysitting every step.

On the numbers, it ties GPT-5.6 Sol Max on the Artificial Analysis composite index — 61 points each — with a large jump on that same DeepSWE coding benchmark. It's live in Cursor, xAI's own Grok Build tool, the API, and OpenRouter, starting around $2 in / $6 out per million tokens.

Gemini 3.7 Flash landed today, roughly three weeks after 3.6 Flash — Google's own product lead called that a "strong intelligence increase" in three weeks, credited to algorithmic improvements, not just more compute. It's better at debugging and shipping deployable code on the first try, live immediately in the API, AI Studio, Antigravity, and now powering Gemini Spark.

Pricing is $0.75 in / $3.75 out per million tokens — introductory pricing through the end of the year, half of what the previous model launched at.

Quick reminder on Meta's Muse Glimmer — the 30-billion-parameter open model from earlier this week — still the local, private, zero-ongoing-cost option if you don't want to run any of today's three through an API. OpenAI shipped a ChatGPT and Codex desktop app for Linux this week, in preview, for technical solo operators stuck using the browser. And Suno's numbers are still worth knowing even though they're not new this week — over 100 million users, 2 million paying subscribers, around $300 million a year in revenue — still the most practical tool out there if you need original audio.

Here's what I'd actually do with three flagship models dropping in two days:

1. Pick one and run it against a real task you already do — don't just read the benchmarks.
2. If cost is your actual constraint, start with DeepSeek V4 Pro before that price hike lands.
3. If you need it live inside a coding tool you already use, Grok 4.6 in Cursor is the path of least friction.
4. If you're doing high-volume, low-complexity work, Gemini 3.7 Flash at half the previous price is the one to route that through.
5. Don't try to run all three. Pick based on the actual constraint you have — cost, integration, or volume — and go deep on one.