Meta Releases Muse Spark 1.3 on 2 September 2026
Meta released Muse Spark 1.3 on 2 September 2026, its fourth Muse Spark model in five months and its most efficient release yet for autonomous software engineering. The model is available through Muse Code, Meta's coding interface, and the Meta Model API at $1.25 per million input tokens and $4.25 per million output tokens — pricing unchanged since July 2026. Muse Spark 1.3 landed in the same week as OpenAI's GPT-6 Astra and Google's Gemini 3.8 Flash, making early September 2026 the densest cluster of frontier model launches since the August wave.
What Changed: Fewer Tool Calls, Fewer Tokens
The headline improvement in Muse Spark 1.3 is not raw capability but operational efficiency. The model completes agentic coding tasks using approximately 20 per cent fewer tool calls and 25 per cent fewer tokens than Muse Spark 1.2. In multi-step autonomous agent workflows — where a model plans, writes code, executes it, reads failure output, and iterates without human checkpoints — fewer tool calls mean lower cost per task and shorter end-to-end completion time. On the Artificial Analysis Intelligence Index, this efficiency advantage puts Muse Spark 1.3 at approximately $0.40 per task at a comparable intelligence tier, against $0.86 for Kimi K3 and $1.18 for GPT-5.5.
The model also handles longer-horizon work better than its predecessor. It can sustain multiple concurrent workflows within a single long thread, reducing the need for developers to restart context between related sub-tasks.
Benchmark Results
Muse Spark 1.3 scores 75.4 on DeepSWE v1.1, the primary benchmark for end-to-end agentic software engineering. On Terminal-Bench 2.1, which measures autonomous task completion in a terminal environment, it scores 88.8. On SWEAtlas CodeBase QnA, which tests whether a model can answer questions accurately about a real codebase's structure and behaviour, the score is 59.4. The Artificial Analysis Intelligence Index places it at 61, level with GPT-5.6 Sol and Grok 4.6.
Long-Context Retrieval
The most significant differentiation over comparable frontier models is at the far end of the context window. On long-context retrieval benchmarks between 512,000 and 1,000,000 tokens, Muse Spark 1.3 holds at 98.1 accuracy, while GPT-5.6 Sol drops to 73.8 at the same range. For teams working with very large codebases, multi-session agent conversations, or extensive regulatory documentation, this is the performance difference that decides which model can reliably complete the task.
Speed and Output Throughput
At approximately 235 tokens per second, Muse Spark 1.3 is the second-fastest frontier model currently available, behind only Google's Gemini 3.8 Flash. For interactive developer workflows where a team member waits for an agent to complete a task before reviewing and directing the next step, throughput at this level makes the difference between a productive flow state and a slow iteration cycle.
Pricing and Context Window
Muse Spark 1.3 maintains the $1.25 per million input token and $4.25 per million output token pricing from Muse Spark 1.2. Against GPT-6 Astra at $10 per million input tokens and Claude Fable 5.1 at the same rate, Muse Spark's pricing is substantially lower at a comparable intelligence tier. The context window is 1,000,000 tokens.
What This Means for Software Teams in India
For development teams in India selecting a frontier model for agentic coding pipelines, Muse Spark 1.3 introduces a cost-per-task argument that changes the calculation for high-volume agent workloads. At $0.40 per task on the Intelligence Index against $1.18 for GPT-5.5 at a comparable capability level, the monthly cost of running autonomous coding, code review, and test generation workflows at scale differs by a factor that becomes substantial once agent usage moves from experimental to routine.
The 20 per cent reduction in tool calls is relevant for teams whose agents interact with external services — version control APIs, CI systems, code review platforms — where each tool call adds latency and potential API quota consumption. Fewer tool calls means faster pipelines and reduced secondary costs beyond token pricing.
The long-context retrieval performance at 512K to 1M tokens matters for Indian product companies managing large monorepos, teams building retrieval-augmented generation pipelines over large knowledge bases, and organisations working with extensive compliance and regulatory documentation that must remain in context across long agent sessions.
The Bottom Line
Meta released Muse Spark 1.3 on 2 September 2026, its fourth Muse Spark model in five months. The model uses approximately 20 per cent fewer tool calls and 25 per cent fewer tokens than Muse Spark 1.2, reducing cost per agentic coding task to approximately $0.40 on the Artificial Analysis Intelligence Index. Benchmark scores: 75.4 on DeepSWE v1.1, 88.8 on Terminal-Bench 2.1, 59.4 on SWEAtlas CodeBase QnA, and 98.1 on long-context retrieval at 512K to 1M tokens — compared with GPT-5.6 Sol's 73.8 at the same range. Generation speed is approximately 235 tokens per second. Pricing is $1.25 per million input tokens and $4.25 per million output tokens. Context window is 1,000,000 tokens. Available through Muse Code and the Meta Model API.
Frequently Asked Questions
What is Meta Muse Spark 1.3 and when was it released?+
Meta Muse Spark 1.3 is an agentic coding model released on 2 September 2026, the company's fourth Muse Spark release in five months. It scores 75.4 on DeepSWE v1.1, 88.8 on Terminal-Bench 2.1, and 59.4 on SWEAtlas CodeBase QnA. The model uses approximately 20 per cent fewer tool calls and 25 per cent fewer tokens than Muse Spark 1.2 to complete the same agentic coding tasks. It supports a 1,000,000-token context window and generates approximately 235 tokens per second. Pricing is $1.25 per million input tokens and $4.25 per million output tokens, unchanged from Muse Spark 1.2. The model is available through Muse Code and the Meta Model API.
How does Muse Spark 1.3 compare to GPT-5.6 Sol on benchmarks?+
On the Artificial Analysis Intelligence Index, Muse Spark 1.3 scores 61, level with GPT-5.6 Sol and Grok 4.6. On DeepSWE v1.1, it scores 75.4. The critical differentiation comes at the far end of the context window: on long-context retrieval benchmarks between 512,000 and 1,000,000 tokens, Muse Spark 1.3 holds at 98.1 accuracy while GPT-5.6 Sol drops to 73.8. On speed, at approximately 235 tokens per second, it is second only to Google's Gemini 3.8 Flash among current frontier models. On cost per task, it achieves approximately $0.40 on the Intelligence Index against $0.86 for Kimi K3 and $1.18 for GPT-5.5 at a comparable intelligence tier.
What does the 20% fewer tool calls improvement mean for agentic coding?+
In agentic coding workflows, each tool call represents an interaction between the model and an external system — writing or reading a file, running a command, calling a CI API, or submitting a pull request. Each tool call adds latency and, when external systems have usage limits or per-call costs, additional expense beyond the model's token pricing. Muse Spark 1.3 completes agentic coding tasks with approximately 20 per cent fewer tool calls and 25 per cent fewer tokens than Muse Spark 1.2, which translates to faster end-to-end task completion, lower pipeline cost, and reduced external API quota consumption for the same work. At scale, this efficiency difference accumulates into substantial cost savings for teams running high volumes of autonomous coding or review tasks.
How is Muse Spark 1.3 priced and where is it available?+
Muse Spark 1.3 is priced at $1.25 per million input tokens and $4.25 per million output tokens — unchanged since Muse Spark 1.2 in July 2026. Against GPT-6 Astra at $10 per million input tokens and Claude Fable 5.1 at the same rate, Muse Spark 1.3 is substantially cheaper at a comparable intelligence tier, placing its cost per task at approximately $0.40 on the Artificial Analysis Intelligence Index. The model supports a 1,000,000-token context window and generates approximately 235 tokens per second. It is available through Muse Code, Meta's coding interface, and the Meta Model API.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.