Morning Edition · Thursday, August 6, 2026Published at 1:47 AM EDT · New York
Meta Ships Muse Code Terminal Agent With Co-Trained Muse Spark 1.2 Model
Meta reports Terminal-Bench 2.1 rising from 76.2 to 82.9 on its own harness. It also introduces a contributor pricing tier at $0.10 per million input tokens in exchange for training rights over user prompts.
Meta released Muse Code, a terminal coding agent, together with Muse Spark 1.2 on 5 August, a direct entry against Anthropic's Claude Code and OpenAI's Codex. Muse Code installs with a single command on macOS or Linux, connects to the Meta Model application programming interface (API), and coordinates persistent background subagents across a repository. Unlike its two main rivals, it ships with no desktop application and runs entirely in the terminal.
The model and the agent were trained together, which is the substantive claim. Meta reports Terminal-Bench 2.1 moving from 76.2 to 82.9, DeepSWE v1.1 from 53.0 to 59.3, and an internal coding benchmark from 68.3 to 70.6 against Muse Spark 1.1. Scoring used the share of tasks solved on the first attempt (pass@1), averaged over five attempts on the 89 tasks in the official Terminal-Bench 2.1 release. The runs took place inside isolated Daytona sandboxes using Meta's own evaluation framework, with each model driven by its vendor's own agent product.
Those are vendor numbers on a vendor harness, and Meta says so, noting that its agent tools and system prompts may not be tuned for third-party models. One analysis points out that no independently verified Terminal-Bench entry exists for the model and that Meta's previous release scored 3.8 points below its own claimed figure. Treat the 82.9 as an upper bound until the public leaderboard reports a result of its own.
The pricing is the more immediately actionable change. Standard rates are unchanged from Muse Spark 1.1 at $1.25 per million input tokens, $0.15 cached input and $4.25 output, with a context window of 1,048,576 tokens. Meta added a contributor tier at $0.10 input and $0.20 output, roughly a twelvefold cut on input and a twentyfold cut on output, in exchange for permission to train future Meta models on submitted prompts and completions. Access runs through the self-serve Meta Model API with no waitlist.
- If true, who benefits
Meta, which converts a benchmark headline and a discounted tier into developer coding trajectories, the input it lacks and cannot buy, at the expense of Anthropic and OpenAI margin on the same workload.
- The nuance
The story omits the comparison that changes the reading: independent write-ups place Muse Spark 1.2 at 82.9 behind Anthropic's Opus 5 at 86.7 and roughly level with GPT-5.6 Terra and Grok 4.5, and every one of those figures comes from a vendor harness rather than the public leaderboard.
An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.
What this means
Meta is using price to acquire training data, not selling capability at a discount. The contributor tier converts developer workloads into a corpus of coding trajectories, the scarcest input for training agentic models, and any team that opts in trades its proprietary codebase context for a lower bill. That pressures Anthropic and OpenAI in two ways at once: on per-token price, where they hold the premium tier, and on data supply, where Meta now has a mechanism to narrow the difference without owning an installed base of coding tools. Enterprises with contractual confidentiality obligations cannot use the cheap tier, so the discount effectively segments the market by legal exposure.
What to watch
- Whether an independent Terminal-Bench 2.1 entry appears for Muse Spark 1.2, since a result materially below 82.9 would repeat the pattern from Meta's last release.
- How many developers accept the contributor tier, because heavy uptake would give Meta a coding-data pipeline that price cuts alone cannot replace.
- Whether Anthropic or OpenAI respond with matching data-for-discount tiers, which would signal that training trajectories, not inference margin, is the binding constraint on coding models.
Observations to monitor, not financial advice.
Synthesized from: Meta AI · Meta AI Research · VentureBeat · Kingy AI
Part of a tracked trend
Frontier Labs Race on AI Coding Capability
Coding is becoming a primary competitive battleground among frontier labs, with incumbents standing up permanent coding teams and investing in new training stages (e.g. midtraining) to match leaders like Anthropic; expect recurring reorganizations, benchmarks, and model releases aimed specifically at code.
More from this edition
- Anthropic Confirms In-House Silicon Team to Design Custom Chips for Claude
- UK AI Security Institute Says Agents Took Unsanctioned Action Against Real Targets in 19 Runs
- Nvidia Releases Alpamayo 2 Super, a 34-Billion-Parameter Driving Model, Under an Open Commercial Licence
- Nvidia Promotes American Chip Manufacturing as Nashville Votes to Seize Land From a Data Center Developer
- Investor Says Safe Superintelligence Plans Its First Model This Month, and the Company Has Not Confirmed It
- New Papers Automate Multimodal Jailbreak Discovery and Map Frontier AI Risk in Critical Infrastructure
- Berlin Police Begin AI Video Analysis at Kottbusser Tor This Month
- MemArena Benchmark Targets the Gap Between Memory Research and On-Device Personal Assistants
- Two Papers Test Whether Language Models Can Formulate Technical Problems, Not Just Solve Them
- Paper Proposes Structural Verification for Long-Horizon Agents That Cannot Be Trusted to Report on Themselves
- Claims of Closed-Loop Self-Improvement in Enzyme Engineering Outrun the Published Evidence
Comments
0No comments yet.