AI Dev Weekly #21: Meta's $0.20 Coding Agent, Qwen 3.8 Max at 2.4T, AWS Ships Kiro Crew
AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.
Four stories this week. Meta launched a terminal coding agent that costs $0.20 per million output tokens if you let them train on your code. Alibaba shipped a 2.4 trillion parameter model that claims Vision Arena #2. AWS open-sourced their internal agent orchestration platform. And the pricing floor for capable AI keeps dropping.
1. Meta Muse Code: $0.20 terminal coding, at the cost of your data
Meta released Muse Code on August 5, a terminal coding agent powered by Muse Spark 1.2. It runs on macOS and Linux, operates in your terminal, and can plan changes, write code, run tests, and make 1,000+ tool calls in a single session.
The pricing model is the story:
| Tier | Input/1M | Output/1M | What Meta gets |
|---|---|---|---|
| Contributor | $0.10 | $0.20 | Trains future models on your prompts and completions |
| Standard | $2.00 | $4.25 | Your data is not used for training |
That contributor tier is roughly 21x cheaper on output than standard. It is 50x cheaper than Claude Sonnet 5’s output ($10) and 125x cheaper than Opus 5 ($25).
The catch is obvious. Your prompts, your code, your completions become Meta training data. For open-source work or personal projects, maybe that is fine. For proprietary codebases, client work, or anything competitive, it is a non-starter.
Benchmarks (via kingy.ai’s verified analysis):
- Terminal-Bench 2.1: 82.9% (vs Opus 5’s 86.7%, Sol’s 88.8%)
- DeepSWE 1.1: 59.3% (third place, behind Anthropic)
Strong on terminal-based agentic coding. Less dominant on isolated code generation without tools. The gap to Opus 5 is small (3.8 points on Terminal-Bench) but the gap on DeepSWE suggests Muse Code works best as a terminal agent, not a pure code generation model.
Muse Spark 1.2 vs 1.1: this is a coding-focused update over July’s 1.1 (which focused on computer-use and multi-app workflows). Same architecture, different training emphasis. 1.2 for code, 1.1 for desktop automation.
My take: The privacy trade-off makes this a split decision. If you work on open-source or non-sensitive code, $0.20/M output is absurdly cheap for 82.9% Terminal-Bench capability. If you work on anything proprietary, you are paying $4.25 (standard tier) which makes it competitive with but not cheaper than Claude Code. The contributor tier will attract hobbyists, students, and open-source maintainers. Enterprise will stay on standard or stick with Claude/Kiro.
2. Qwen 3.8 Max: 2.4 trillion parameters, but verify before you trust
Alibaba launched Qwen 3.8 Max on August 3. The headline specs:
- 2.4T total parameters, 95B active (Sparse MoE)
- 1M context window
- Multimodal (text + vision)
- Open weights promised “next week”
- Text Arena #5, Vision Arena #2, Frontend Code Arena #4
- 16-day autonomous coding (built “oh-my-cli” from scratch without human intervention)
This is Alibaba putting itself back in the frontier conversation. The 2.4T parameter count makes it the largest model from any Chinese lab (surpassing Kimi K3’s 2.8T total but with fewer active parameters: 95B vs K3’s estimated 200B).
The skepticism: Multiple independent analysts (edenai.co, emergent.sh) note that Qwen 3.8 Max has “no independent benchmarks, model card, or license yet.” The Arena rankings are from Alibaba’s own submissions. The 16-day autonomous coding claim is an internal demonstration, not a reproducible benchmark.
Compare with Kimi K3, which shipped open weights, published pricing ($3/$15), and has independently verified Terminal-Bench scores (88.3%). Qwen 3.8 Max has a preview endpoint with credit-based pricing and vendor-reported numbers.
My take: Wait for the open weights (promised this week) and independent benchmarks before making decisions. The Vision Arena #2 ranking is intriguing if it holds up. But “announced” and “independently verified” are different things in this industry, and Chinese frontier models have a history of impressive demos followed by slower real-world availability than the headlines suggest. See our Qwen 3.8 Max vs GPT-5.6 Sol and vs Kimi K3 comparisons for the detailed breakdown.
3. Kiro Crew: AWS makes AI agents persistent
AWS launched Kiro Crew on August 4. This is not a coding assistant, it is a persistent workspace that keeps AI agents running 24/7 across sessions.
What it does:
- Agents persist across sessions with memory, lessons, and skills
- Scheduled jobs via cron (morning digests, PR monitoring, flaky test detection)
- Webhook triggers (Slack messages, GitHub events, CI failures start agent work)
- Multi-agent coordination with isolated contexts
- Purpose-built Apps (DevFleets, Issue Radar, Task Runner)
- Security: OS-level sandbox, denied-by-default commands, signed audit logs
Origin: Internal Amazon project called “MeshClaw.” 39,000+ Amazon builders used it in less than 6 months before open-sourcing. 500+ contributors.
How it differs from Kiro IDE: Kiro IDE is an AI code editor for interactive sessions. Kiro Crew is for work that happens when you are not there. Think of it as the difference between a coding assistant and a coding teammate that works nights.
How it compares: We published a Kiro Crew vs Hermes Agent comparison. Kiro Crew is for teams that need coordinated multi-agent workflows. Hermes Agent (225K GitHub stars, Nous Research) is for individuals who want a single agent that learns and grows over time. Different tools, different philosophies.
My take: The internal Amazon adoption (39K builders in 6 months) is the strongest signal. This is not a demo or a research project, it is a tool that Amazon’s own developers chose to use at scale. The cron/webhook model is the right abstraction for “AI that works while you sleep.” But it is tightly coupled to the Kiro/AWS ecosystem, so evaluate Hermes Agent if you want model-agnostic flexibility.
4. The pricing floor in one chart
Updated this week. The cheapest capable model at each tier:
| Tier | Model | Input/1M | Output/1M | Change |
|---|---|---|---|---|
| Ultra-budget | DeepSeek V4 Flash | $0.14 | $0.28 | unchanged |
| Budget (with data sharing) | Muse Spark 1.2 contributor | $0.10 | $0.20 | NEW |
| Budget (private) | GPT-5.6 Luna | $0.20 | $1.20 | -80% (Jul 30) |
| Mid-tier | GPT-5.6 Terra | $2.00 | $12.00 | -20% (Jul 30) |
| Frontier coding | Claude Sonnet 5 | $2.00 | $10.00 | unchanged |
| Frontier reasoning | Claude Opus 5 | $5.00 | $25.00 | unchanged |
| Peak benchmark | GPT-5.6 Sol | $5.00 | $30.00 | unchanged |
The floor dropped twice in one week. Luna’s 80% cut on July 30 made $0.20 the new “cheap but capable” price. Meta’s contributor tier on August 5 matched it, but with a data-sharing requirement.
For most developers, the practical choice is now:
- Luna at $0.20/$1.20 for private, high-volume work
- Muse Spark contributor at $0.10/$0.20 for open-source or non-sensitive work
- DeepSeek V4 Flash at $0.14/$0.28 for the absolute cheapest option (but smaller context)
The “is AI too expensive?” question is over. The question is now “which cheap model is good enough for my use case?” See our full pricing comparison.
That is it for this week. Next Thursday: Qwen 3.8 Max weights (if they actually ship), Muse Code’s first real-world reception, and whatever else breaks.
Want this in your inbox? Subscribe to AI Dev Weekly.
FAQ
What is the cheapest good coding model right now?
GPT-5.6 Luna at $0.20/$1.20 if you want privacy. Muse Spark 1.2 contributor at $0.10/$0.20 if you are comfortable sharing your code with Meta. Both score above 80% on Terminal-Bench.
Is Qwen 3.8 Max better than Kimi K3?
Cannot say definitively. Qwen 3.8 Max has vendor-reported Arena rankings but no independent benchmarks. Kimi K3 has independently verified Terminal-Bench (88.3%) and open weights available today. Wait for Qwen weights and independent testing.
Should I switch from Claude Code to Muse Code?
Only if you use the contributor tier ($0.20) and your code is not sensitive. At the standard tier ($4.25), Muse Code is more expensive than Claude Code with similar capabilities. Claude Code still leads on DeepSWE benchmarks.
Is Kiro Crew free?
The Crew workspace is open-source and free. You pay for model API calls and Kiro CLI usage (free tier available). No additional cost for the Crew orchestration itself.