Meta Shipped Muse Spark 1.2 and a Coding Agent — It's Good, It's Closed, and It's Still Not Opus 5
Meta Shipped Muse Spark 1.2 and a Coding Agent — It's Good, It's Closed, and It's Still Not Opus 5
Hi guys, Mr. Technology here.
Meta dropped Muse Code (beta) and Muse Spark 1.2 today as a paired release — terminal coding agent plus the model behind it, co-trained together so the harness and the weights are actually optimized for each other. That part is genuinely novel. The pricing is competitive. The model is closed, requires a Facebook login (yes, that Facebook), and on every public coding benchmark it lands exactly one slot behind Opus 5. Meta is now a real third option in the agentic-coding war, not a marketing stunt. It's just not the top option.
What You Need to Know
>
- Meta released Muse Code (beta) and Muse Spark 1.2 on August 5, 2026. The agent is a Mac/Linux terminal tool; the model is API-accessible globally through the Meta Model API (now in public preview). - Muse Spark 1.2 has 1,048,576-token context (1M), pricing $1.25 / $0.15 cached / $4.25 per million tokens on the standard tier, or $0.10 / $0.002 / $0.20 on a contributor tier that lets Meta train on your data. - On Terminal-Bench 2.1: 82.9% — second place, behind Opus 5's 86.7% (Claude Code), ahead of GPT-5.6 Terra 81.8% (Codex) and Grok 4.5 81.6% (Grok Build). - On DeepSWE 1.1: 59.3% — third place, behind Opus 5 (65.0%) and GPT-5.6 Terra (64.8%). - On Meta Internal Coding Bench: 70.6% — second place, behind Opus 5's 79.4%, ahead of Muse Spark 1.1 (68.3%) and GPT-5.6 Terra (65.4%). - On the KDA kernel optimization case study (1,000+ tool calls, up to 24 hours): +68.7% speedup vs the FLA Triton baseline — fourth place, behind Opus 5 (+74.0%), GPT-5.6 Sol (+71.2%), Opus 4.8 (+69.6%). - Spark 1.1 was released less than three weeks ago, on July 16. This is, very visibly, a re-launch.
The Architecture, Briefly
Meta isn't publishing parameter counts or pre-training details for Spark 1.2 — the model card is vendor charts and not much else. What they did publish is the interesting part:
- Co-training with the agent. Spark 1.2 was trained with rejection-sampled Muse Code trajectories and explicit recipe optimization for goals, context compaction, and subagent orchestration. Model and harness shaped for each other in the same RL loop.
- Long-horizon training. Whole-repository generation, end-to-end projects, "auto-research." The post calls out planning, goal conditioning, and context compaction — exactly the failure modes of long-running coding agents.
- Self-improvement loop. Spark 1.1 generated challenging coding environments and instruction-following templates; solutions were graded, producing a scalable dataset for 1.2. The same trick quietly powering most 2026 reasoning-model gains, applied to coding.
- Persistent background agents. Muse Code keeps async background agents alive for the whole session instead of spawning them per-task. Lower latency, less redundant context gathering, less steering. The runtime uses a local event log for replay-exact, restart-safe resumption after crashes.
Real decisions, not a vibes launch. The trade-off is the model is closed and you log in through Meta's identity stack — which is the part the Hacker News thread got loud about.
The Benchmarks, Read Honestly
Meta published charts, not a methodology write-up. The chart numbers, though, are clear:
| Benchmark | Opus 5 (max, Claude Code) | Muse Spark 1.2 (Muse Code) | GPT-5.6 Terra (max, Codex) | Grok 4.5 (high, Grok Build) | Gemini 3.6 Flash (high, Antigravity CLI) | Muse Spark 1.1 (mini-swe-agent) |
|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 86.7% | 82.9% | 81.8% | 81.6% | 78.9% | 76.2% |
| DeepSWE 1.1 | 65.0% | 59.3% | 64.8% | 56.6% | 40.0% | 53.0% |
| Meta Internal Coding | 79.4% | 70.6% | 65.4% | — | 63.9% | 68.3% |
| KDA kernel speedup | +74.0% | +68.7% | +65.1% (Terra) / +71.2% (Sol) | — | +62.5% | — |
Things to notice:
1. Spark 1.2 vs 1.1 is a real jump, mostly from co-training. Terminal-Bench +6.7 (76.2 → 82.9). DeepSWE +6.3 (53.0 → 59.3). Meta Internal +2.3 (68.3 → 70.6) — the smallest gain on Meta's own benchmark, which is what you'd expect if the lift came from harness integration rather than raw model improvement.
2. It doesn't beat Opus 5 anywhere. Not one benchmark, not one case study. The closest it gets is Terminal-Bench, 3.8 points behind. If you're picking a coding model on absolute quality in early August 2026, Opus 5 is still the answer.
3. On DeepSWE it's behind GPT-5.6 Terra by 5.5 points. Real code patches against real repos (TypeScript, Go, Python, JS, Rust) with hand-written verifiers and regression checks — not a vibes benchmark.
4. On the kernel-optimization case study, the ranking matters. Meta picked this as a headline: 1,000+ tool calls over 24 hours writing, compiling, profiling, iteratively improving GPU kernels. Opus 5 wins (+74.0%). GPT-5.6 Sol wins the model class it was designed to win (+71.2%). Opus 4.8 — the previous generation — still beats Spark 1.2. The HN commenter who noted "their model is worse" was reading the chart correctly.
Pricing: Aggressive, With a Catch
Two tiers on the Meta Model API:
- Standard: $1.25 input / $0.15 cached input / $4.25 output per million tokens. No training on your data.
- Contributor: $0.10 / $0.002 / $0.20 per million tokens. Meta may train on your prompts and outputs.
Comparison on the standard tier:
- Claude Opus 5: $5 / $25 per million → Spark 1.2 is 4× cheaper on input, 6× cheaper on output
- GPT-5.6 (Sol): ~$3 / $15 per million → Spark 1.2 is 2.4× cheaper on input, 3.5× cheaper on output
- Grok 4.5: $1.50 / $6 per million → Spark 1.2 is roughly even on input, 1.4× cheaper on output
- DeepSeek V4 Flash: $0.10 / $0.40 per million → Spark 1.2 is more expensive, but closed-source and Western
The contributor tier puts Spark 1.2 at effectively DeepSeek V4 Flash pricing — the actual price-war shot across the bow. The data-retention split is also unusual: most major labs either train on everything by default or don't train on anything (Claude with the no-retention API setting). Meta's "give us your data and we'll charge you 12× less" structure is closer to how social networks have always priced privacy.
What's Actually Interesting
Two real things and one sus thing.
The co-training is the right idea. Most coding agents are a generic model plus a wrapper. The wrapper has opinions about when to compact, when to spawn subagents. The model has opinions about the same things. They fight. Meta's bet — train the model in the harness so they don't fight — is the same bet that made Claude Code / Opus 5 good and Codex / GPT-5.6 good. The Spark 1.2 → 1.1 jump (+6.7 on Terminal-Bench) is mostly that bet paying off. Every coding-agent lab will copy this within six months.
1M context is the new floor. Not because 1M tokens is useful for a single coding task — it usually isn't — but because long-horizon loops accumulate state, and state plus compaction plus a 1M headroom is what makes a 24-hour run feasible. Opus 5, GPT-5.6, Grok 4.5, and now Spark 1.2 are in the same ballpark. Labs shipping below 500K will feel it.
Sus: the launch timing. Spark 1.1 was July 16. Spark 1.2 is August 5. Twenty days. Kimi K3 shipped July 31 and absolutely ate the open-weights coding-model news cycle — the segment Spark 1.1 was supposed to compete in. Meta is doing what every frontier lab does when a launch gets overshadowed: ship a do-over. The new benchmarks conveniently drop Kimi K3 from the comparison even though it was on the Spark 1.1 chart.
The Take
Picking a coding agent today:
1. Best absolute quality: Opus 5 + Claude Code. Not close. 2. Best raw price/performance on a closed API: Muse Spark 1.2 + Muse Code, especially with aggressive cached-input routing. 3. Best open-weights story you can self-host: Kimi K3 and DeepSeek V4 Flash. Spark 1.2 doesn't address this segment at all. 4. Best cheap-and-Western: Muse Spark 1.2 contributor tier, at the cost of letting Meta train on your prompts.
Meta has done the hard part: real model, real agent, real benchmarks, real API. What they haven't done is cross the line from "very good second tier" to "top tier." On every chart in the launch post, the bar above them is Opus 5. If Meta thinks this is worth a re-launch three weeks after Spark 1.1, they will keep iterating until one of those charts shows a number above Opus 5. That's the goal post. We'll see if they hit it.
— Mr. Technology
Model: Muse Spark 1.2 (co-trained with Muse Code, the agent) · Lab: Meta Superintelligence Labs · Release date: August 5, 2026 · Context: 1,048,576 tokens (1M) · Pricing (per million tokens): Standard $1.25 input / $0.15 cached / $4.25 output · Contributor $0.10 / $0.002 / $0.20 (data used for training) · API: Meta Model API (public preview) · OpenRouter · Benchmarks: Terminal-Bench 2.1 82.9% (2nd) · DeepSWE 1.1 59.3% (3rd) · Meta Internal Coding Bench 70.6% (2nd) · KDA kernel optimization +68.7% (4th) · Closed: Yes. Login required. Vendor benchmarks only.