Meta Muse Spark Is Fast for Small Agent Tasks — Should Your Agency Use It?

Published August 15, 2026Updated August 30, 2026By ABD Legacy LLC
AI models / model routing

A well-known operator's hands-on test

On August 6, 2026, Julian Goldie (@JulianGoldieSEO) posted a hands-on test of Meta's Muse Spark inside Hermes Agent. His verdict: "Big models still win for complex work. But inside Hermes Agent, Muse Spark is incredibly fast for smaller tasks, making your AI team more efficient." Source: X post, Aug 6, 2026.

That is exactly the model-selection question agencies face every week: which model should run which task? Goldie's answer — fast, cheap model for high-volume small work, frontier model for the hard stuff — is a routing strategy, not just a model review.

What's verified vs. what's still open

Verified: Muse Spark was tested inside Hermes Agent by a prominent AI/SEO operator on Aug 6, 2026. His two claims are on record: very fast for smaller agent tasks; bigger models still win on complex work. Pricing is also on record: Muse Spark 1.2 is Meta's natively multimodal reasoning model at $1.25/M input and $4.25/M output, with a $0.10/$0.20 contributor tier when you allow Meta to use submitted data (Meta docs; Simon Willison, Aug 5, 2026).

Not proven yet: there is no independent speed or latency benchmark for Muse Spark — Artificial Analysis lists Speed: N/A — and no published small-task evals specific to the model. "Fast for small tasks" is a practitioner report, not a benchmark. Treat it as dated attribution, not measured throughput.

What the pricing table says about routing

ModelIntelligence index*Price in/out ($/M)AA cost/taskSpeed (tok/s)
Claude Opus 5 (max)63—$2.3454
Claude Sonnet 5 (PERMANENT, Aug 10 2026)***—2.00 / 10.00——
GPT-5.6 Sol (max) (promo through Nov 21, 2026)****614.00 / 20.00$0.8663
Muse Spark 1.2 (xhigh)571.25 / 4.25$0.40N/A
Gemini 3.8 Flash (new, Sept 2 2026)**—0.75 / 3.75——
Gemini 3.7 Flash (new, Aug 13 2026)**—0.75 / 3.75——
Gemini 3.6 Flash521.50 / 7.50$0.56213
DeepSeek V4 Flash 0731 (pre-hike rate)*52—$0.03106

*Artificial Analysis Intelligence Index (9 evals), Aug 2026. Cost-per-task figures are reference points, not quotes — no independent speed benchmark exists for Muse Spark yet, so size pilot workloads on your own data.

**Gemini 3.8 Flash launched Sept 2, 2026 at the same introductory rate Google set for 3.7 Flash: $0.75/1M input and $3.75/1M output through December 31, 2026, then $1.50/$7.50. Gemini 3.7 Flash launched Aug 13, 2026 at that intro rate — exactly half of Gemini 3.6 Flash's launch price — with Google price-matching 3.6 Flash to the same intro rate through Dec 31, 2026. Artificial Analysis had not published an Intelligence Index, cost-per-task, or speed figure for 3.8 Flash as of launch day; Google-reported DeepSWE v1.1 and HLE-Verified figures had not been independently verified (9to5Google; Google blog, 3.7 Flash; Gemini API pricing docs).

***Claude Sonnet 5 (Anthropic) API pricing is PERMANENT: $2.00 per 1M input tokens / $10.00 per 1M output tokens, made permanent by Anthropic on Aug 10, 2026. The previously scheduled Sept 1, 2026 increase to $3/$15 was CANCELLED — there is no scheduled price change pending. Cache: $0.20 per 1M cached input; cache write $2.50 (5m) / $4 (1h) per 1M. API-pricing change only; subscription prices unchanged (Anthropic announcement; Claude Platform pricing docs).

****GPT-5.6 Sol (OpenAI) API pricing was cut Aug 21, 2026 to $4.00 per 1M input / $20.00 per 1M output tokens (cached input $0.50 → $0.40), a 20%/33% reduction. The discount is promotional through at least Nov 21, 2026 — treat the $4/$20 rate as temporary, with the price expected (not guaranteed) to revert to $5/$30 after the window. Applies to the API plus eligible ChatGPT Work and Codex credits; ChatGPT Pro/Plus/Business subscriptions unchanged (OpenAI developer docs; AWS Bedrock). AA cost/task figure recomputed from the $1.23 pre-cut reference by applying the -20%/-33% price change to a 70/30 output/input cost split (AA's published blend); exact mix varies by workload.

Update — September 2, 2026

Google launches Gemini 3.8 Flash — third Flash in six weeks, same intro price

On September 2, 2026, Google released Gemini 3.8 Flash, its “best reasoning and coding model yet,” three weeks after 3.7 Flash and the third Flash family release in six weeks. Introductory pricing holds at $0.75/1M input and $3.75/1M output tokens through December 31, 2026, stepping to $1.50/$7.50 after — the same intro rate as 3.7 Flash (Ars Technica; 9to5Google).

What this means for the table above: the Gemini column now has two rows at the same intro price — 3.8 Flash above 3.7 Flash — and Google now recommends Flash for software engineering and autonomous agent workloads. Google reports 3.8 Flash at the top of the DeepSWE v1.1 leaderboard at a fraction of the cost of larger frontier models (unverified independently at launch), so for agencies routing small-to-mid agent work on Gemini, the workhorse tier just got its second refresh in a month. Google's cadence (no frontier Pro since early 2026) says plan on Flash refreshes, not a Pro flagship. Read our full Gemini 3.8 Flash vs Claude Fable 5.1 vs GPT-5.6 Sol analysis for the comparison table, benchmark caveats, and scoping specs.

Update — August 21, 2026

OpenAI cuts GPT-5.6 Sol API pricing — $4/$20 promotional through Nov 21, 2026

On August 21, 2026, OpenAI dropped GPT-5.6 Sol API and credit pricing by over 20%: input $5.00 → $4.00, cached input $0.50 → $0.40, output $30.00 → $20.00 per 1M tokens. The discount is promotional through at least November 21, 2026, applies to the API plus eligible ChatGPT Work and Codex credits, and does not change ChatGPT Pro/Plus/Business subscriptions (OpenAI developer docs; AWS Bedrock; OpenAI community announcement).

What this means for the table above: the GPT-5.6 Sol row now shows $4.00 / $20.00 with a recomputed cost/task of $0.86 (down from the $1.23 pre-cut reference). Sol remains the premium frontier tier — still well above Sonnet 5's permanent $2/$10 and Grok 4.6's $2/$6 — so the routing guidance on this page is unchanged: cheap workhorse for high-volume small tasks, frontier model for complex work. But agencies quoting multi-month builds against Sol should note the price is expected (not guaranteed) to revert to $5/$30 after November 21, 2026 — price the promo window into the quote.

Update — August 30, 2026

Claude Code weekly limits change September 14: +25% permanent baseline, net −17% vs today

On August 29, 2026, Anthropic announced that starting September 14, 2026 it is permanently raising Claude Code's standard weekly usage limits by 25% over the pre-promotion baseline for Pro, Max, Team, and seat-based Enterprise plans — and the temporary 50% weekly boost (active since May 13, 2026, extended through the summer) expires the same day (@ClaudeDevs, Aug 29, 2026; BleepingComputer). Anthropic confirmed the net effect itself: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code." The canonical illustration: baseline 100% → today 150% → September 14 125%.

What this means for the routing/cost math on this page: the change applies to Pro, Max (5x and 20x), Team (standard and premium seats), and seat-based Enterprise plans — Free and consumption-based Enterprise seats are excluded — and covers Claude Code only (CLI, IDE extensions, desktop app, web). 5-hour session limits, Claude chat, Claude Cowork, and Console API pay-per-token pricing are unaffected. For agencies running Claude agents at scale, the weekly pool is shared across Claude Code and the Claude apps, there is no published numeric cap for any plan, and /usage in the CLI is the only ground truth — so re-check your monthly math against the 125% post-change baseline and treat the remaining +50% as time-boxed runway through September 13. See Claude Code limits drop 17% on Sept 14 — what agencies should do for the full breakdown.

Update — August 13, 2026

Google launches Gemini 3.7 Flash — the new coding/agent workhorse

On August 13, 2026, Google released Gemini 3.7 Flash, its “most intelligent workhorse model yet for coding and agents,” just three weeks after Gemini 3.6 Flash. Introductory pricing is $0.75/1M input and $3.75/1M output tokens through December 31, 2026 — exactly half of Gemini 3.6 Flash's launch price — before returning to $1.50/$7.50 on January 1, 2027 (Google blog; official Gemini API pricing).

What this means for the table above: 3.7 Flash is Google's answer to the routing question this page is about — a cheap, fast workhorse for high-volume agent and coding tasks, priced below Muse Spark at launch. Google reports big coding/agent gains over 3.6 Flash: FrontierCode 1.1 Main 43.6% vs 34.4%, DeepSWE v1.1 65.3% vs 49.0%, and WebDev Arena Elo 1588 vs 1538. It also price-matched 3.6 Flash to the same $0.75/$3.75 intro rate through year-end, so both Flash tiers are promo-priced into 2027. If your agency routes small-to-mid agent work on Gemini, re-run the routing math now — the workhorse tier just got cheaper. Read our full Gemini 3.7 Flash analysis for the pricing cliff, benchmark table, and scoping specs.

Update — August 10, 2026

Claude Sonnet 5's $2/$10 API pricing is now permanent

Anthropic made Claude Sonnet 5's API pricing permanent on Aug 10, 2026: $2.00 per 1M input tokens and $10.00 per 1M output tokens. The previously scheduled Sept 1, 2026 increase to $3/$15 was CANCELLED — there is no scheduled price change pending (Anthropic announcement; Claude Platform pricing docs). Cache pricing: $0.20 per 1M cached input; cache write $2.50 (5m) / $4 (1h) per 1M. This is an API-pricing change only — subscription prices are unchanged.

What this means for the table above: Sonnet 5's input price sits at open-weight parity (Qwen 3.8 Max lists $2/$6), while its output price is well below premium frontier estimates (GPT-5.6 Sol $4/$20 — promotional through at least Nov 21, 2026 — Artificial Analysis). For routing decisions, Sonnet 5 is a hosted frontier API with no open-weight/self-host upside — agencies that quote multi-month builds against Sonnet 5 no longer need to price in a September repricing event.

Update — August 8, 2026

DeepSeek has announced a significant API price increase

DeepSeek has announced it plans to raise the overall pricing for its API services "in the near future," with a "significant increase" expected — new rates, the percentage, and the effective date have not yet been disclosed. The warning sits on the official DeepSeek pricing page (api-docs.deepseek.com/quick_start/pricing) and was reported on August 6, 2026 by The Next Web ("DeepSeek warns of a 'significant' price rise, reversing its cheap-AI pitch").

What this means for the table above: the DeepSeek V4 Flash 0731 row shows the currently published rate ($0.14/M input, $0.28/M output; ~$0.03 per task on Artificial Analysis) — treat it as provisional, pre-hike pricing. The "DeepSeek is the cheapest option" reading of this table may not survive the increase. Re-run your routing math against the official schedule once the new rates are published, and check the DeepSeek changelog for the effective date (api-docs.deepseek.com/updates).

Independent testing from April 2026 (Ritesh Khanna) found Muse Spark won vision and analysis tasks against Claude Opus 4.6, GPT-5.4, Gemini 3.1, and Grok 4.2 — but finished 4th of 5 on a one-shot complex code task. "Meta crushed vision and analysis but face-planted on code."

What this means for your agency

Route high-volume, structured, small tasks — triage, classification, metadata extraction, short copy — to a fast, cheap model like Muse Spark. Keep complex multi-file coding and long-horizon work on frontier models. Cost-per-task drops when small tasks run on a cheap model and expensive runs are reserved for work that needs them.

Price AI-assisted work with the AI agency cost calculator

Estimate Your Cost Per Task →

Or browse the findaiagency.com directory for agencies that route models deliberately.

Frequently asked questions

What did Julian Goldie report about Muse Spark?

In a hands-on test posted August 6, 2026, Julian Goldie (@JulianGoldieSEO) reported that inside Hermes Agent, Meta's Muse Spark is "incredibly fast for smaller tasks, making your AI team more efficient," while "big models still win for complex work." This is a practitioner report, not a published benchmark.

Is Muse Spark actually faster for small tasks?

Not independently proven yet. There is no public output-speed or latency benchmark for Muse Spark — Artificial Analysis lists Speed as N/A. "Fast for small tasks" is a dated practitioner report, not a benchmark.

What does Muse Spark cost?

Muse Spark 1.2 is priced at $1.25 per million input tokens and $4.25 per million output tokens, with a $0.10/$0.20 contributor tier when you allow Meta to use submitted data.

What should my agency route to Muse Spark?

Small, structured, high-volume tasks — triage, classification, metadata extraction, and short copy — where speed and price matter. Keep complex multi-file coding and long-horizon research on frontier models.

How does Muse Spark's cost per task compare to frontier models?

Artificial Analysis (August 2026) lists Muse Spark 1.2 at about $0.40 per task versus $2.34 for Claude Opus 5 and $0.03 for DeepSeek V4 Flash — but DeepSeek has announced a significant API price increase with new rates and timing not yet disclosed, so treat the DeepSeek figure as provisional pre-hike pricing. No independent speed benchmark exists for Muse Spark yet, so size pilot workloads on your own data.

Is Claude Sonnet 5's $2/$10 pricing permanent or scheduled to increase?

Permanent. Anthropic made Claude Sonnet 5's API pricing permanent on Aug 10, 2026: $2.00 per 1M input tokens and $10.00 per 1M output tokens (cache hit $0.20 per 1M input; cache write $2.50 for 5m / $4 for 1h per 1M). The previously scheduled Sept 1, 2026 increase to $3/$15 was CANCELLED — there is no scheduled price change pending. This is an API-pricing change only; subscription prices are unchanged. Input price sits at open-weight parity (Qwen 3.8 Max $2/$6), while output is well below premium frontier estimates (GPT-5.6 Sol $4/$20 — promotional through at least Nov 21, 2026 — Artificial Analysis).

Is Claude Code's 50% weekly usage-limit boost permanent?

No — and the change is now dated. Anthropic announced August 29, 2026 that the temporary 50% weekly-limit boost (active since May 13, 2026, previously extended through August 31) ends on September 14, 2026, when a permanent 25% increase over the pre-promotion baseline takes effect for Pro, Max, Team, and seat-based Enterprise plans. Because 125% replaces 150%, paid users see a net 17% reduction in weekly Claude Code capacity versus today. It covers Claude Code only (CLI, IDE extensions, desktop app, web); 5-hour session limits, Claude chat, Claude Cowork, and Console API pay-per-token pricing are unaffected. Free and consumption-based Enterprise seats are excluded.

Related reading

Sources