Muse Spark 1.2 and Muse Code: benchmarks, pricing and comparison with Claude Opus 5, GPT-5.6, Kimi K3 and Grok 4.5

The model was not released with open weights. Artificial Analysis classifies Muse Spark 1.2 as a proprietary model, and its parameter count has not been publicly disclosed.^[3]

The most interesting part of the release is not the model name itself. Meta trained Muse Spark 1.2 together with Muse Code and optimized the model for a real agentic environment: planning, context compaction, subagents, tool use and long-running tasks that span entire repositories.^[1]

The results are strong, but they require careful interpretation.

In Meta's official evaluation:

• Muse Spark 1.2 + Muse Code reaches 82.9% on Terminal-Bench 2.1,
• 59.3% on DeepSWE 1.1,
• 70.6% on Meta Internal Coding Bench.^[1][2]

That does not mean Muse Spark 1.2 is the best coding model on the market. In the same tables, Claude Opus 5 + Claude Code ranks first in all three comparisons. Meta also explicitly notes that the comparison is not a perfect model-to-model test: each model ran with a different agent, and the evaluation setup may have been better tuned for Muse Code than for competing tools.^[2]

Artificial Analysis provides a more independent view. On the current model card checked on August 7, Muse Spark 1.2 at xhigh scores 57 on Artificial Analysis Intelligence Index v4.1.1. For comparison, Claude Opus 5 scores 61, GPT-5.6 Sol 59, Kimi K3 57 and GPT-5.6 Terra 55.^{[3][7][8][9][10]}

Shortest conclusion: Muse Spark 1.2 does not unconditionally beat Opus 5 or GPT-5.6 Sol, but it is exceptionally competitive on price. It looks strongest as an agentic coding model, especially when paired with Muse Code and when the cost of long sessions matters.

All data in this article was verified against Meta materials, the official evaluation methodology, Artificial Analysis and independent press coverage. Status as of August 7, 2026.

TL;DR

Question	Verified answer
When was Muse Spark 1.2 released?	August 5, 2026
Who created the model?	Meta
What is Muse Code?	A terminal-based coding agent in beta
Does Muse Spark 1.2 have open weights?	No, the model is proprietary
Do we know the parameter count?	No, Meta has not publicly disclosed it
Context	1M tokens
Modalities	text and image input, text output
Standard input price	$1.25 / 1M tokens
Cached input price	$0.15 / 1M tokens
Output price	$4.25 / 1M tokens
Terminal-Bench 2.1 according to Meta	82.9%
DeepSWE 1.1 according to Meta	59.3%
Meta Internal Coding Bench	70.6%
Current AA Intelligence Index v4.1.1	57
AA Index in the launch-day article	54
Are Meta's results independent?	No, some are vendor-run evaluations
Is Artificial Analysis independent of Meta?	Yes, although Meta provided pre-release access for benchmarking
Strongest area	agentic coding and capability-to-price ratio
Main limitation	it does not win every benchmark and some tests use different agent harnesses

1. What exactly did Meta release?

The launch includes two related products:

Muse Spark 1.2
model

Muse Code
terminal coding agent

Meta describes Muse Spark 1.2 as a coding-focused update to Muse Spark 1.1. The company increased the amount of compute used during training on programming tasks and expanded the diversity of training environments.^[1]

Muse Code is an agent that can:

• plan repository changes,
• write code,
• run tools,
• validate results,
• coordinate several long-running subagents,
• maintain progress across long sessions.^[1]

Meta documents installation of the official client on macOS or Linux as follows:

curl -fsSL https://dev.meta.ai/install.sh | bash

Muse Spark 1.2 is available through both Muse Code and Meta Model API. Meta describes current availability as broadly expanded globally.^[1]

2. Muse Code is not just a simple wrapper around the model

For coding agents, a model benchmark and a product benchmark are not the same thing.

Muse Code adds its own execution layer.

Async Background Agents

The primary agent can use specialized subagents that continue working in the background. They are not created only for a single step and can remain active throughout the session.^[1]

This is intended to reduce:

• repeatedly gathering the same information,
• latency,
• the need for manual steering,
• repeated repository analysis.

Append-only event log

Muse Code stores a local history of:

• model calls,
• tool usage,
• approvals,
• edits.^[1]

Meta describes the runtime as replay-exact and restart-safe. After a failure, the agent can resume from the log rather than starting the task from the beginning.

That is a meaningful feature for multi-hour coding tasks.

Built-in skills

The initial version includes commands such as:

/plan
/grill
/goal

/plan prepares an approval-gated plan, /grill critically evaluates the plan, and /goal keeps the work directed toward a defined objective.^[1]

3. The model and agent were trained together

This is one of the most important facts for interpreting the benchmarks.

Meta did not simply attach a new model to an existing CLI. The company says it co-trained Muse Spark 1.2 and Muse Code.^[1]

Training included, among other things:

• trajectories from the agent harness,
• optimizations for goal-based work,
• compaction,
• subagents,
• the integrated Muse Code toolset.^[1]

The model was also trained on long-horizon tasks such as:

• generating complete repositories,
• large end-to-end projects,
• auto-research,
• sequencing work through planning,
• maintaining direction through goal conditioning,
• context compaction.^[1]

This helps explain why the proper comparison unit is sometimes:

Muse Spark 1.2 + Muse Code

rather than only:

Muse Spark 1.2

4. Self-improvement using Muse Spark 1.1

Meta used Muse Spark 1.1 to build training data for version 1.2.

According to the official description, the earlier model:

1. generated difficult coding environments,
2. prepared templates for complex instructions,
3. evaluated candidate solutions,
4. helped create a scalable dataset for Muse Spark 1.2.^[1]

Meta calls this a self-improvement loop.

This does not mean the model “trained itself.” It was still a controlled training pipeline designed by Meta, with the previous model serving as one component for generating and assessing data.

5. Official Terminal-Bench 2.1

Terminal-Bench 2.1 evaluates agents performing tasks in a terminal environment.

Meta used all 89 tasks from the official 2.1 release. Each run took place in an isolated Daytona sandbox, and an executable verifier evaluated the final container state. The score is the average pass@1 over five attempts.^[2]

Results published by Meta

Model + agent	Effort	Terminal-Bench 2.1
Claude Opus 5 + Claude Code	max	86.7%
Muse Spark 1.2 + Muse Code	xhigh	82.9%
GPT-5.6 Terra + Codex	max	81.8%
Grok 4.5 + Grok Build	high	81.6%
Gemini 3.6 Flash + Antigravity CLI	high	78.9%
Muse Spark 1.1 + mini-swe-agent	xhigh	76.2%	[1][2]

Muse Spark 1.2 improves on its predecessor by 6.7 percentage points.

At the same time it:

• trails Opus 5 by 3.8 pp,
• leads GPT-5.6 Terra by 1.1 pp,
• leads Grok 4.5 by 1.3 pp.

The key caveat

This is not a pure benchmark of the underlying models alone.

The compared systems are:

model
+
agent
+
tools
+
system prompt
+
runtime

Meta uses:

• Muse Code for Spark 1.2,
• Claude Code for Opus 5,
• Codex for GPT-5.6,
• Grok Build for Grok 4.5,
• Antigravity CLI for Gemini.^[2]

Meta itself notes that the configuration and system prompts may not have been optimally tuned for the closed competing models.^[2]

Therefore the correct conclusion is:

In Meta's setup, Muse Spark 1.2 + Muse Code achieved 82.9%.

Not:

Muse Spark 1.2 is universally better than GPT-5.6 Terra.

6. DeepSWE 1.1

DeepSWE v1.1 contains 113 tasks from 91 repositories across five languages:

• TypeScript,
• Go,
• Python,
• JavaScript,
• Rust.^[2]

Each task has a manually prepared functional verifier and regression tests.

During rollout and grading, external internet access is disabled. The model endpoint remains available.^[2]

Meta results

Model + agent	DeepSWE 1.1
Claude Opus 5 + Claude Code	65.0%
GPT-5.6 Terra + Codex	64.8%
Muse Spark 1.2 + Muse Code	59.3%
Grok 4.5 + Grok Build	56.6%
Muse Spark 1.1 + mini-swe-agent	53.0%
Gemini 3.6 Flash + Antigravity CLI	40.0%	[1][2]

Muse Spark 1.2 improves on 1.1 by 6.3 pp.

Here, however, the gap to the leaders is larger:

Opus 5       65.0%
GPT-5.6      64.8%
Muse 1.2     59.3%

It is a strong result, but not first place.

Additional methodology caveat

The official DeepSWE leaderboard uses mini-swe-agent for every model.

Meta's evaluation instead used a separate agent product for each model. The company explicitly says the result is not harness-identical to the official leaderboard.^[2]

That is an important qualification that should appear next to the 59.3% figure.

7. Meta Internal Coding Bench

Meta Internal Coding Bench contains 440 tasks sourced from Meta's internal codebase and real pull requests.^[2]

It covers:

• bug fixes,
• new features,
• refactoring,
• cleanup,
• other software-engineering tasks.

Internet access is disabled. Solutions are compiled and checked with unit tests. Meta runs two attempts per task.^[2]

Results

Model	Meta Internal Coding Bench
Claude Opus 5	79.4%
Muse Spark 1.2	70.6%
Muse Spark 1.1	68.3%
GPT-5.6 Terra	65.4%
Gemini 3.6 Flash	63.9%	[1][2]

Muse Spark 1.2 improves on its predecessor by 2.3 pp.

Why does this benchmark have less external value?

Not because it must be wrong.

The issue is reproducibility.

The dataset:

• is internal,
• comes from Meta's private codebase,
• is not a public benchmark that outsiders can independently reproduce.

The result should therefore be treated as an additional vendor signal, not as independent proof of superiority.

8. Independent check: Artificial Analysis

Artificial Analysis received access to Muse Spark 1.2 from Meta before public release and ran its own evaluation suite.^[4]

This matters because it separates:

vendor benchmark

from:

independent benchmark organization

Important score change after launch

Artificial Analysis' August 5 article reported:

Muse Spark 1.2 xhigh
Artificial Analysis Intelligence Index: 54

The current model card, after the index moved to v4.1.1, shows:

Muse Spark 1.2 xhigh
Artificial Analysis Intelligence Index: 57

^[3][4]

The number 54 should therefore not be copied into a current ranking without a date.

Benchmarks evolve. They change through:

• index versions,
• grading,
• evaluation sets,
• methodology fixes.

A production article should always identify the snapshot.

9. Current Artificial Analysis comparison

Model cards checked on August 7, 2026:

Model	AA Intelligence Index
Claude Opus 5, max	61
GPT-5.6 Sol, max	59
Muse Spark 1.2, xhigh	57
Kimi K3, max	57
GPT-5.6 Terra, max	55
Grok 4.5, high	54*	[3][7][8][9][10][11]

\* The Grok 4.5 score comes from Artificial Analysis' published report on that model. Values should be treated as snapshots of the relevant benchmark versions, not permanent ranking positions.

What does this tell us?

Muse Spark 1.2:

• does not currently catch Opus 5 in the overall index,
• remains 2 points behind GPT-5.6 Sol,
• is level with Kimi K3 on the current AA model card,
• leads GPT-5.6 Terra in the overall index,
• sits above Grok 4.5's launch-era result.

But Artificial Analysis Intelligence Index is not a coding-only benchmark.

The current v4.1.1 combines nine evaluations, including:

• GDPval-AA v2,
• τ³-Banking,
• Terminal-Bench v2.1,
• SciCode,
• Humanity's Last Exam,
• GPQA Diamond,
• CritPt,
• AA-Omniscience,
• AA-LCR.^[3]

The score of 57 should therefore be read as a broader signal of intelligence and agentic capability.

10. GDPval-AA v2: the biggest improvement over Spark 1.1

In Artificial Analysis' launch snapshot, GDPval-AA v2 rose from:

Muse Spark 1.1: 1371 Elo
Muse Spark 1.2: 1631 Elo

an increase of 260 Elo points.^[4]

GDPval-AA v2 tests realistic knowledge-work tasks.

It contains 220 tasks from:

• 44 occupations,
• 9 major sectors of the U.S. economy.^[2]

Models create deliverables such as:

• documents,
• spreadsheets,
• presentations,
• diagrams,
• reports.

Artificial Analysis uses its own Stirrup agentic harness with shell and web-browsing access, and outputs are compared pairwise by an LLM judge.^[2]

Snapshot from August 5

Artificial Analysis reported:

Model	GDPval-AA v2 Elo
Claude Opus 5	1852
GPT-5.6 Sol	1730
Kimi K3	1685
Muse Spark 1.2	1631
Claude Opus 4.8	1588	[4]

This is one of the strongest pieces of evidence that 1.2 improved beyond code generation alone.

The GDPval ranking is updated over time, so these numbers should not be mixed with later leaderboard snapshots.

11. Terminal-Bench according to Artificial Analysis

Meta reports:

82.9%

for Muse Spark 1.2 + Muse Code.^[1]

Artificial Analysis reported at launch in its own harness:

80%

compared with:

78%

for Muse Spark 1.1.^[4]

These results are not contradictory.

They come from:

• different harnesses,
• different agent configurations,
• different execution procedures.

This is a good example of why a benchmark result should never be published without methodology.

12. Hallucinations: the number looks better, but read the other half of the table

Artificial Analysis' launch snapshot showed an improvement in AA-Omniscience:

18 -> 22

and a lower hallucination rate:

38% -> 28%

^[4]

At first glance, that looks unambiguously positive.

However, at the same time:

attempt rate: 82% -> 67%
accuracy:     41% -> 38%

^[4]

The model more often declined to answer when uncertain.

AA-Omniscience does not penalize abstention. The lower hallucination rate is therefore partly a result of greater willingness to abstain.

Correct interpretation:

Muse Spark 1.2 was less likely to present an incorrect answer with confidence, but it also attempted fewer answers.

Incorrect interpretation:

Muse Spark 1.2 became 10 percentage points more accurate.

The data does not support that claim.

13. Scientific reasoning did not improve uniformly

In Artificial Analysis' launch measurements:

Benchmark	Spark 1.1	Spark 1.2	Change
CritPt	15%	18%	+3 pp
SciCode	58%	56%	-2 pp
Humanity's Last Exam	45%	44%	-1 pp	[4]

This matters because it shows the nature of the update.

Muse Spark 1.2 is not simply:

Spark 1.1
+ a few percent everywhere

The largest launch-day gains were concentrated in:

• agentic coding,
• tool use,
• professional agentic tasks.

14. API pricing: one of Muse Spark 1.2's biggest strengths

Meta's standard Muse Spark 1.2 pricing:

Tokens	Price per 1M
Cached input	$0.15
Input	$1.25
Output	$4.25	[3][5][6]

Input/output pricing is unchanged from Muse Spark 1.1.

Comparison of public API prices

Model	Input / 1M	Output / 1M
Muse Spark 1.2	$1.25	$4.25
GPT-5.6 Terra	$2.50	$15.00
Kimi K3	$3.00	$15.00
Claude Opus 5	$5.00	$25.00
GPT-5.6 Sol	$5.00	$30.00	[3][7][8][9][10]

This compares token prices, not the cost of finishing a task.

A cheaper model per million tokens can:

• generate more tokens,
• make more tool calls,
• require more iterations,
• make an error that needs manual repair.

The more useful business metric is therefore:

cost per correctly completed task

not only:

price per 1M tokens

15. Cost per task according to Artificial Analysis' launch measurement

In the August 5 snapshot, Artificial Analysis reported the weighted average cost of an Intelligence Index task:

Model / configuration	Task cost
Grok 4.5 high	$0.37
GPT-5.6 Sol medium	$0.39
Muse Spark 1.2 xhigh	$0.40
GPT-5.6 Terra max	$0.51
Kimi K3 max	$0.86
GPT-5.5 xhigh	$1.18	[4]

That is a very strong cost result for Muse Spark 1.2.

Compared with Spark 1.1, however, cost rose from:

$0.29
to
$0.40

Artificial Analysis links this to greater token usage:

• about 53% more input tokens,
• about 36% more output tokens,

especially in GDPval-AA v2.^[4]

In other words, version 1.2 is:

better
but
more token-intensive

The unit API price did not change, but the agent does more work.

16. Muse Spark 1.2 vs Claude Opus 5

Where does Opus 5 win?

In Meta's data:

Benchmark	Opus 5	Muse 1.2
Terminal-Bench 2.1	86.7%	82.9%
DeepSWE 1.1	65.0%	59.3%
Meta Internal Coding Bench	79.4%	70.6%	[1]

In the current Artificial Analysis Intelligence Index:

Opus 5:     61
Muse 1.2:   57

^[3][7]

There is no evidence in the available data to call Muse Spark 1.2 the better model overall.

Where does Muse have an advantage?

Primarily in token pricing:

Muse Spark 1.2
$1.25 / $4.25

Claude Opus 5
$5 / $25

^[3][7]

For agents performing thousands of operations, that difference can matter significantly.

Conclusion

Opus 5 is the better choice when:

• maximum success rate matters more than token cost,
• the task is difficult and the cost of failure is high,
• the team already works in Claude Code.

Muse Spark 1.2 makes more sense when:

• task volume is high,
• the agent works for a long time on a repository,
• token cost matters heavily,
• a few benchmark points do not justify a much more expensive API.

17. Muse Spark 1.2 vs GPT-5.6 Terra

This is one of the most interesting comparisons.

Meta Terminal-Bench

Muse 1.2 + Muse Code: 82.9%
Terra + Codex:         81.8%

Meta DeepSWE

Muse 1.2 + Muse Code: 59.3%
Terra + Codex:         64.8%

Meta Internal Coding Bench

Muse 1.2:  70.6%
Terra:     65.4%

^[1]

The leader changes depending on the benchmark.

Artificial Analysis

Current overall index:

Muse Spark 1.2 xhigh: 57
GPT-5.6 Terra max:    55

^[3][9]

Price

Muse:
$1.25 / $4.25

Terra:
$2.50 / $15

^[3][9]

This does not mean Muse automatically wins. Terra is clearly stronger on DeepSWE in Meta's comparison.

The most defensible interpretation is:

Muse Spark 1.2 is exceptionally competitive with GPT-5.6 Terra on price and some agentic tasks, but the winner depends on task type and harness.

18. Muse Spark 1.2 vs GPT-5.6 Sol

GPT-5.6 Sol remains stronger in Artificial Analysis' broader index:

GPT-5.6 Sol max: 59
Muse Spark 1.2:  57

^[3][8]

Sol costs:

$5 / 1M input
$30 / 1M output

versus:

$1.25 / 1M input
$4.25 / 1M output

for Muse.^[3][8]

That is a large difference for:

• long reasoning traces,
• subagents,
• iterative debugging,
• repository-wide refactoring,
• multi-hour agent runs.

If the extra 2 points in the overall index do not translate into a higher correct-completion rate in a company's real tasks, the more expensive model may not be more economical.

19. Muse Spark 1.2 vs Kimi K3

The current Artificial Analysis cards show:

Muse Spark 1.2 xhigh: 57
Kimi K3 max:          57

^[3][10]

That does not mean they have identical capabilities.

The key difference

Kimi K3:

• has publicly available weights,
• can be self-hosted,
• has 2.8T total parameters and 104B active parameters,
• uses its own Kimi K3 License.^[10]

Muse Spark 1.2:

• is proprietary,
• runs through Meta Model API,
• has no publicly disclosed parameter count.^[3]

Official API pricing

Muse:
$1.25 input
$4.25 output
$0.15 cached input

Kimi K3:
$3 input
$15 output
$0.30 cached input

^[3][10]

If self-hosting is not required, Muse is much cheaper in nominal API pricing.

If an organization needs:

• its own infrastructure,
• control over model weights,
• custom model modifications,

Kimi K3 offers something Muse Spark 1.2 currently does not.

20. Muse Spark 1.2 vs Grok 4.5

In Meta's Terminal-Bench:

Muse Spark 1.2 + Muse Code: 82.9%
Grok 4.5 + Grok Build:      81.6%

On DeepSWE:

Muse Spark 1.2: 59.3%
Grok 4.5:        56.6%

^[1]

In Artificial Analysis' launch-era Intelligence Index, Grok 4.5 scored 54, similar to Muse 1.2's launch snapshot before the index update.^[4][11]

Artificial Analysis also listed Grok 4.5 high as one of the few models with a lower task cost in the same quality cluster:

Grok 4.5 high:  $0.37
Muse 1.2 xhigh: $0.40

^[4]

This is another example of why cost-quality trade-offs in 2026 are no longer a simple rule of:

most expensive model
=
most economical model

21. Kernel optimization: more than 1,000 tool calls and up to 24 hours of work

Meta ran an interesting long-horizon experiment.

Muse Code and the models were asked to iteratively optimize GPU kernels through:

• more than 1,000 tool calls,
• up to 24 hours,
• writing code,
• compiling,
• profiling,
• repeated optimization.^[1]

The test covered KDA and MLA on NVIDIA Hopper.

KDA

The baseline was an FLA implementation in Triton. The models were not allowed to directly import external kernel libraries such as FLA.^[1]

In the final speedup-vs-baseline chart, Meta reported, among others:

Model	Speedup vs baseline
Claude Opus 5	+74.0%
GPT-5.6 Sol	+71.2%
Claude Opus 4.8	+69.6%
Muse Spark 1.2	+68.7%
GPT-5.6 Terra	+65.1%
Gemini 3.6 Flash	+62.5%	[1]

Muse Spark 1.2 did not win this experiment.

That matters because the official Meta material does not present its own model as the leader in every category.

What does this experiment not prove?

It does not prove that:

• Opus 5 always optimizes code by 74%,
• Muse always produces a 68.7% speedup,
• the results transfer to every kernel or GPU.

It is a case study for a specific task, baseline, harness and hardware setup.

22. Long context: 1 million tokens

Artificial Analysis confirms for Muse Spark 1.2:

1M token context window

and:

• text input,
• image input,
• text output.^[3]

A million tokens can matter for:

• large repositories,
• monorepos,
• documentation,
• long agent history,
• multi-file analysis.

It does not mean you should dump an entire repository into a prompt.

Maximum capacity does not guarantee:

• finding every dependency,
• correct prioritization,
• resistance to malicious text in the repository,
• identical quality at the beginning and end of the context.

Muse Code therefore also uses context compaction, rather than relying only on ever-larger prompts.^[1]

23. Why is the current AA score 57 when launch articles may say 54?

This is not an error to hide.

Artificial Analysis published a score of 54 on August 5.^[4]

The current Muse Spark 1.2 model card checked on August 7 shows 57 and identifies the current Artificial Analysis Intelligence Index as v4.1.1.^[3]

In practice, benchmark citations should preserve:

model
+
effort
+
benchmark version
+
date

For example:

Muse Spark 1.2 (xhigh)
AA Intelligence Index v4.1.1
57
status: August 7, 2026

This is much better than writing:

Muse Spark 1.2 scores 57

without context.

24. Standard tier vs Contributor tier: an important privacy difference

Standard Muse Spark 1.2 pricing is:

input:        $1.25 / 1M
cached input: $0.15 / 1M
output:       $4.25 / 1M

^[3][5][6]

Meta also offers a heavily discounted Contributor tier.

Independent press coverage describing the official offer reports that the lower price is tied to agreeing that user activity may be used to improve Meta products.^[6]

For enterprise use, this is an important distinction.

Before using Contributor tier for code that is:

• private,
• covered by an NDA,
• containing trade secrets,
• owned by a client,

check the exact data-processing terms applying to the account and endpoint.

Do not select a cheaper endpoint based only on price.

25. Is Muse Spark 1.2 the best model for coding?

Based on verified data: there is no basis for that claim.

It does not lead every Meta benchmark

Opus 5 leads in:

• Terminal-Bench 2.1,
• DeepSWE 1.1,
• Meta Internal Coding Bench.^[1]

GPT-5.6 Terra also leads Muse on DeepSWE.

But it is exceptionally strong on price

Muse Spark 1.2 has much lower output pricing than:

• Opus 5,
• GPT-5.6 Sol,
• GPT-5.6 Terra,
• Kimi K3.^{[3][7][8][9][10]}

Best description

The most defensible conclusion is:

Muse Spark 1.2 is one of the most interesting 2026 models in terms of agentic-coding capability per dollar, but Opus 5 remains stronger in the available quality comparisons, and the final choice should come from your own POC.

26. Which model should you choose?

Choose Muse Spark 1.2 if:

• the cost of long sessions matters a lot,
• you work with large repositories,
• you want to use Muse Code,
• the agent must work for a long time without manual steering,
• you need persistent subagents,
• nominal output pricing must stay low,
• near-frontier performance is enough for the workflow.

Choose Claude Opus 5 if:

• maximum success rate is the priority,
• the cost of an error is higher than token cost,
• you already use Claude Code,
• tasks are exceptionally hard,
• your own POC confirms a higher completion rate.

Choose GPT-5.6 Terra if:

• you use the Codex ecosystem,
• DeepSWE-like tasks are central,
• you need a strong coding model at lower cost than Sol,
• OpenAI tools are already part of the pipeline.

Choose GPT-5.6 Sol if:

• you need higher general capability than Muse,
• API cost matters less,
• the workload uses broader reasoning beyond coding.

Choose Kimi K3 if:

• you need open weights,
• you plan to self-host,
• you need custom model modifications,
• you can accept higher official API pricing.

Consider Grok 4.5 if:

• your own tests show strong agentic performance,
• cost per completed task is the key metric,
• you use Grok Build or a compatible harness.

27. How to run a fair POC

Do not test models with five prompts.

Prepare at least:

50-200 real tasks

from your own repository.

Categories

• bug fix,
• new feature,
• refactoring,
• tests,
• code review,
• dependency migration,
• performance optimization,
• log analysis,
• frontend from a screenshot,
• repo-wide change,
• technical documentation.

Measure for every task

success / failure
time
cost
token count
tool-call count
manual interventions
regressions

Most important metric

Not:

which model wrote the nicest answer?

But:

which model most often delivered a correct, merge-ready result
at an acceptable cost and time?

28. How to compare agents, not just models

For:

• Muse Code,
• Claude Code,
• Codex,
• Grok Build,

the model does not operate in isolation.

The agent decides:

• which files to read,
• when to run tests,
• when to grep,
• whether to use a subagent,
• how to manage context,
• how many iterations to perform,
• when to stop.

It is therefore useful to run two tests.

Test A: model in a shared harness

same agent
same tools
same system prompt
same limits

This helps compare the underlying models.

Test B: end-to-end product

Muse + Muse Code
Opus + Claude Code
GPT + Codex

This answers the practical question:

Which solution is best for the team?

The two tests measure different things.

29. Production checklist for Muse Spark 1.2

Benchmarks

• Benchmark version is recorded.
• Score date is recorded.
• Reasoning effort is recorded.
• Agent name is recorded.
• Meta Terminal-Bench is not mixed with the AA result.
• Meta Internal Coding Bench is not treated as an independent benchmark.
• Results are verified on the team's own repository.

Cost

• Actual input usage is measured.
• Cached input is measured.
• Output is measured.
• Reasoning cost is measured.
• Tool-call count is measured.
• Cost is calculated per completed task.
• At least three models are compared.
• A per-agent spending limit is defined.

Repository

• The agent has the minimum required permissions.
• Secrets are not stored in easily accessible files.
• Production access requires approval.
• Merge requires review.
• Tests run automatically.
• The agent cannot bypass branch protection.
• Changes and tool calls are logged.

Data

• Standard tier terms are verified.
• Contributor tier terms are verified.
• Client code does not go to the wrong endpoint.
• Company policy permits the selected service.
• Prompt and log retention rules are defined.
• Personal data is properly protected.

Quality

• The agent runs tests after changes.
• The final diff is reviewed.
• Regressions are measured.
• Long tasks have checkpoints.
• Failure recovery is tested.
• Large monorepos are tested.
• Multi-hour tasks are tested.

30. POLPROG verdict

Muse Spark 1.2 is a more significant release than the 1.2 version number might suggest.

Meta built the model and Muse Code as a connected system, focusing training on:

• large repositories,
• long-running tasks,
• tools,
• subagents,
• planning,
• automated verification.

The results confirm a real improvement over Muse Spark 1.1.

Strongest facts

Terminal-Bench 2.1
76.2% -> 82.9% in Meta's evaluation

DeepSWE 1.1
53.0% -> 59.3% in Meta's evaluation

GDPval-AA v2
1371 -> 1631 Elo in Artificial Analysis' launch snapshot

^[1][4]

But it is still not the leader everywhere

Opus 5 beats Muse Spark 1.2 on all three coding benchmarks shown by Meta.

In the current Artificial Analysis snapshot:

Opus 5      61
GPT-5.6 Sol 59
Muse 1.2    57
Kimi K3     57
Terra       55

^{[3][7][8][9][10]}

So why is Muse Spark 1.2 interesting?

Because it costs:

$1.25 / 1M input
$4.25 / 1M output
$0.15 / 1M cached input

^[3]

That makes the following question less useful:

Which model has the highest benchmark score?

A better question is:

Which model delivers the highest number of correctly completed tasks per $100 of budget?

For many teams, the answer may be different from the model ranked first in one table.

Muse Spark 1.2 currently looks like a very strong candidate for large-scale agentic coding with controlled cost.

It is not an unconditional winner.

It is a model that clearly deserves a place in a POC alongside:

• Claude Opus 5,
• GPT-5.6 Terra,
• GPT-5.6 Sol,
• Kimi K3,
• Grok 4.5.