Five verified moves this week, and a team of one can act on each before Monday. Meta stopped watching the coding-agent war and joined it — with a price that asks for your code in return. OpenAI put two deprecation dates on the calendar that will break something you built if you ignore them. Anthropic opened a free usage window that closes on the 19th. The money moved, again, to the layer around the models. And the open-weight coding backend is about to get another serious option. Every item below is dated and sourced; each carries the one line that changes what you do next.
1. Meta shipped a terminal coding agent — and priced it for your data#
On August 5, Meta released Muse Code (beta), its first terminal coding agent, powered by the new Muse Spark 1.2 model (MarkTechPost; Yahoo Finance). It has the now-standard shape — persistent async background agents, parallel sub-agents, git-worktree isolation, a crash-safe event log — putting it head-to-head with Claude Code, OpenAI Codex, and Google's Antigravity CLI. Muse Spark 1.2 posts 77.4 on SWE-Bench Verified and about 54 on the Artificial Analysis Intelligence Index (Artificial Analysis) — a real coding-focused step up from 1.1, though on every benchmark Meta itself chose to show, it trails the current Claude and Gemini leaders in the low 80s.
What it means: The headline isn't the score, it's the price with a string attached. Standard pricing is about $1.25/$4.25 per 1M input/output, but a contributor tier drops input to roughly $0.10/1M — in exchange for letting Meta train on your sessions and code. For a solo builder that's a genuinely cheap background-agent backend for throwaway or open work. For anything private or client-owned, the discount is the product: read the contributor terms before you pipe a repo through it. We put the whole field side by side in Muse Code vs Claude Code vs Codex for solo founders, and the raw cost math still lives in what it actually costs to run a coding agent this month.
2. OpenAI set two deprecation deadlines you have to clear#
OpenAI put two dates on the calendar. The Atlas browser agent stops working August 9 as OpenAI folds browser-based agentic capability into ChatGPT and Codex. And on August 26, OpenAI retires o3 from ChatGPT (after a 90-day sunset) and removes the Assistants API from the platform (OpenAI deprecations; ChatGPT Atlas).
What it means: These are dependency deadlines, not news. If you built a workflow on Atlas, move it into ChatGPT or Codex this weekend — the 9th is days away. If you built on the Assistants API, migrate to the Responses and Conversations APIs using OpenAI's migration guide before the 26th; the shape changes but the features carry over. This is the same clear-your-calendar drill we ran for the two August deadlines that raise your agent bill — put both on it now, while the fix is an afternoon, not an outage.
3. Claude Code opened a free usage window — and hardened agent secrets#
Anthropic's August Claude Code build shipped a Focus view (Ctrl+Alt+F, which hides tool activity behind an expandable per-turn summary), sandbox credential masking — a mask mode on Linux and WSL where a sandboxed command reads a sentinel copy of a secret and the proxy substitutes the real value only on egress — and inference-hook DLP in beta for Enterprise. The lever most builders can use today: Anthropic extended a 50% weekly usage boost through August 19, applied automatically to eligible Pro and Max accounts (Releasebot).
What it means: Two concrete actions. First, turn on credential masking for any agent run that touches real secrets — it's the cheapest reduction in blast radius you'll get this week, and it pairs with the broader permission-hardening moves the coding CLIs shipped this month and the case for containing a coding agent's shell before it can run an RCE. Second, spend the usage window: schedule the batch refactors, migrations, or long agent runs you've been rationing before the boost lapses on the 19th.
4. The money moved to the agent-ops layer, again#
Sapiom raised a $35M Series A on August 5 — led by Dragonfly, with Accel, Gradient, Coinbase Ventures, and Anthropic among the backers — to help builders ship, run, and scale AI agents. It brings the company to $50M total, 11 months after founding (Sapiom).
What it means: It's the third signal in a month that capital is pooling in the layer around the models — identity, logging, rollout, governance — not the models themselves, echoing Zenity's $125M for the agent-security layer and the broader vertical-and-infrastructure funding pattern. You don't need to buy any of it to get the lesson: the moment an agent touches production, its scoped credentials and an action log are the difference between an incident you can explain and one you can't. Solve that yourself this week even if you never buy the platform.
5. The open-weight coding backend is about to get another contender#
Watch for Zhipu's GLM-5.5 — a 1T+-parameter, coding-agent-focused open-weight flagship that leaks put on an August release, though Zhipu hasn't confirmed a model card, benchmark, or endpoint as of this week (kie.ai). It would land on top of an already-crowded open-weight shelf: Moonshot's Kimi K3 shipped in July, and Alibaba previewed Qwen3.8-Max (a claimed 2.4T multimodal model, weights promised but not yet released).
What it means: Treat this as a standing instruction, not a single event: keep one open-weight backend wired into your model router so the next credible option — GLM-5.5, or whatever follows — is a config change, not a migration. The existence of a strong open-weight coding model is the leverage that keeps every closed vendor's pricing honest. We walked through the wiring in pointing your coding agent at Kimi K3 through OpenRouter; the same router slot is where GLM-5.5 goes the day it ships.
The throughline: the coding agent became a five-way market this week (item 1), the plumbing under it kept professionalizing and getting funded (items 3 and 4), and the cost floor kept falling — sometimes in tokens, sometimes in your data (items 1 and 5). For a founder, the job this week is small and concrete: clear OpenAI's two deadlines before they clear you, turn on credential masking, spend the free usage window, and keep an open-weight backend one config line away. The decision underneath all of it — which terminal coding agent to actually run — is the piece we wrote next.
