Gemini 3.8 Flash vs Muse Spark 1.3: Which affordable AI is better

Google and Meta released Gemini 3.8 Flash and Muse Spark 1.3 on the same day, September 2, 2026, and both target the same buyer: teams that need near-frontier results without frontier pricing. Neither model wins outright. Muse Spark 1.3 posts higher independent intelligence scores with the lowest verified cost per task at its level, plus near-perfect long-context retrieval and the stronger coding peak. Gemini 3.8 Flash counters with lower promotional per-token pricing, leading finance and legal agent scores, and broader multimodal input. The right pick follows the workload, the cache pattern, and two pricing caveats that change the arithmetic.

Release status and access differences

Both models are generally available now, through different channels. Gemini 3.8 Flash ships across the Google ecosystem: API, AI Studio with a free tinkering tier, and the Gemini app for Pro and Ultra subscribers. A separate 3.8 Flash Cyber variant for vulnerability discovery and patching stays behind the Fairwind Program for trusted defenders, governments, and infrastructure operators. Muse Spark 1.3 ships through Muse Code and the Meta Model API under the model ID muse-spark-1-3, with an OpenAI-compatible API and OpenRouter availability. Its top max reasoning mode remains in limited partner preview pending additional safety testing, so the generally available tier at launch is xhigh.

The spec sheets align. Both list a 1M-token context window (1,048,576 on both). Gemini caps output at 64K tokens with text, image, audio, video, and PDF input and text output, plus configurable low, medium, and high thinking levels defaulting to medium. Spark accepts text, image, video, and document input with configurable reasoning effort. Knowledge cutoff is March 2026 for some Gemini domains and January 2025 for others, a split Google discloses in the model card. Meta has published no single cutoff date for Spark 1.3 in the launch materials surveyed.

Agentic coding and software engineering

Coding is the closest contest in this comparison, with each side holding one benchmark family.

Spark holds the higher peak on long-horizon engineering. Meta reports 75.4% on DeepSWE v1.1 for Spark 1.3 max, against 73.0% for GPT-5.6 Sol max, 74.0% for Claude Opus 5 max, and 55.0% for Spark 1.2 xhigh. Google reports 73.7% on the same benchmark for Gemini 3.8 Flash, against 74.0% for Opus 5, 72.7% for Sol, and 65.3% for 3.7 Flash. Both vendors place Opus 5 near 74%, so the two challengers sit within about two points of each other and of Opus 5, with Spark about two points ahead of Flash. Spark adds 59.4% on SWEAtlas CodeBase QnA for repository understanding, against 53.5% for Sol and 52.7% for Opus 5, and Meta engineers report about 20% fewer tool calls and 25% fewer tokens than 1.2 on internal coding comparisons. An independent hands-on test of three medium repository tasks through the Spark API passed all checks for $0.04 in total token spend.

Flash holds the edge on the older terminal benchmark and concedes the newer one. Google reports 89.4% on Terminal-Bench 2.1 for Flash against 89.1% for Opus 5, while Meta reports 88.8% for Spark 1.3 max, tied with Sol and 2.1 points ahead of Opus 5 on Meta's setup. On the current and far harder Terminal-Bench 4.0, Google reports 19.1% for Flash against 51.8% for Opus 5, a wide deficit Google published itself. Meta has published no Terminal-Bench 4.0 figure for Spark 1.3, so that row stays open.

repository-level engineering and codebase QnA favor Spark on current figures. Well-trodden terminal tasks are a draw, while the hardest agentic coding benchmark is a documented weak spot for Flash and an unknown for Spark.

Computer use across desktop and browser

Agents that operate real software favor Spark, with both trailing the Claude tier.

Meta reports 66.9% on OSWorld 2.0 for Spark 1.3 max, against 68.3% for Opus 5 max, 62.7% for Sol max, and 47.6% for Spark 1.2 xhigh. Google reports 59.0% for Flash on the same benchmark family against 75.4% for Opus 5 and 50.6% for 3.7 Flash. Vendor harnesses differ, so treat the Spark-to-Flash gap of about eight points as a directional signal and confirm exact sizing on team-owned reruns. On agentic browsing, Meta reports 89.4% on DeepSearchQA for Spark against 93.0% for Sol and 90.4% for Opus 5, so browsing retrieval is competitive without leading.

desktop-automation teams pick Spark between these two. Teams needing top-tier computer use look above this price band entirely.

Knowledge work, finance, and legal analysis

Professional knowledge work splits by benchmark origin: Spark leads the general knowledge-work index, Flash leads finance and legal agent tests.

Spark scored 1,754 Elo on GDPval-AA v2 knowledge work in max mode and 1,709 in xhigh, against 1,824 for Opus 5 max, 1,710 for Sol, and 1,615 for Spark 1.2. Flash scored 1,545 on the same evaluation against 1,482 for 3.7 Flash and 1,824 for Opus 5, a gap of roughly 160 to 210 Elo behind Spark's two modes. Artificial Analysis reflects the same ordering: Spark xhigh at Intelligence Index 61 and max at 62 in limited preview, Flash high at 59.

On finance and legal agent benchmarks, only Flash has published figures in this price band. Google reports 61.4% on Vals Finance Agent v2 for Flash against 58.6% for Opus 5 and 53.8% for Sol, and 10.0% on Harvey's Legal Agent Benchmark against 6.7% for Opus 5 and 2.5% for Sol. Absolute legal scores stay low everywhere, and 10% represents the current state of the art on that test, still far from production-ready autonomy. Spark's adjacent professional signal is JobBench tool use at 64.9%, against 65.7% for Opus 5 and 45.4% for Sol, plus 49.4% on AutomationBench end-to-end workflows against 50.3% for Opus 5 and 46.7% for Sol.

general analyst-desk work and multi-step professional tool use favor Spark. Finance and legal agent pipelines favor Flash on the only published figures in this price band, subject to reruns on team-owned evals.

Scientific reasoning and lab work

Flash holds the broader published science record here. Google reports 86.2% on LABBench2 lab science against 84.2% for Opus 5, 56.5% on the human-difficult split of BioMysteryBench against 49.4% for Opus 5 and 44.7% for Sol, and 86.2% on CharXiv chart reasoning against 83.7% for Opus 5. Humanity's Last Exam verified lands at 54.9% for Flash, ahead of Sol at 54.5%, Opus 5 at 54.4%, and 3.7 Flash at 53.6%. Artificial Analysis attributes Spark's four-point Index climb largely to agentic-work and scientific gains, and Meta reports 57.8% on an instruction-following index plus 98-plus long-context retrieval scores that underpin research loops, but Meta has published no matched LABBench or BioMystery figures for Spark 1.3.

lab-science, chart-heavy, and bio-reasoning pipelines pick Flash on published evidence. Spark's science case rests on composite-index gains, with no matched head-to-head science tables published, so research teams should test both on their own corpora.

Long context, video, and multimodal reach

Spark's widest technical lead is long-context retrieval. Meta reports 98.5% on MRCR 256K to 512K and 98.1% on 512K to 1M for Spark 1.3 max, against 91.5% and 73.8% for Sol and 66.3% and 55.5% for Spark 1.2 xhigh. No matched MRCR figures exist for Flash in the surveyed materials. Flash answers with long-video understanding: 87.8% on LVBench in agentic video mode against 75.4% for Opus 5. Input breadth also favors Flash, which accepts audio alongside text, image, video, and PDF, while Spark covers text, image, video, and documents.

retrieval over 500K-plus contexts points to Spark. Video-centric and audio-inclusive pipelines point to Flash.

Safety and behavioral signals

Both vendors claim safety progress, with different evidence. Google reports a 5.5% attack success rate for Flash on the Gray Swan prompt-injection benchmark, against 4.8% for Opus 5, 51.8% to 60.1% for Grok 4.6, Kimi K3, and DeepSeek V4 Pro. Flash did not reach any track or critical capability level under Google's Frontier Safety Framework assessment. Meta reports stronger adversarial-input resistance, better prompt-injection handling, and improved calibration on irreversible actions for Spark 1.3, plus a working style that asks clarifying questions on ambiguous requirements and confirms before consequential actions. Those are vendor-reported behavioral claims without matched third-party tables at publication time.

The Cyber variants sit outside this comparison by design: Flash Cyber posts 86.2% on CyberGym vulnerability discovery and 47.2% pass@1 on CWE-Bench patching near a leading frontier model at 47.8%, with Chrome Security reporting 2.6 times the correct patches of larger commercial models. It remains gated behind Fairwind.

injection-resistance numbers favor Flash where published figures exist. Teams governing consequential agent actions weigh Spark's confirmation behavior in their own trials.

Cost, speed, and the two pricing caveats

Sticker pricing favors Flash, verified task cost slightly favors Spark, and two caveats qualify both.

Flash lists $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, doubling to $1.50 and $7.50 on January 1, 2027. Batch and Flex inference halve the promo rates, Priority raises them to $1.35 and $6.75, cached input reads cost $0.075, and cache storage runs $0.50 per million tokens per hour. Thinking tokens bill at the output rate with reasoning on by default, so metered cost routinely exceeds naive visible-output estimates; Google exposes the count as total_thought_tokens.

Spark standard pricing holds at $1.25 per million input, $0.15 cached input, and $4.25 output, unchanged from 1.2. A separate contributor endpoint at roughly $0.10 input, $0.002 cached, and $0.20 output trades steep discounts for Meta training on the traffic. Artificial Analysis measured Spark xhigh at $0.55 per Intelligence Index task, the lowest figure at Index 59 or above, against $0.58 for Flash high, $0.63 for Sol xhigh, $0.68 for GLM-5.3 max, and $0.94 to $1.23 for Grok, Sol max, and Opus peers. Spark task cost rose from $0.40 on 1.2 on 57% higher input tokens per agentic task, and community testing pegs Spark output volume near 3 times its predecessor, so verbosity offsets per-token price on long traces.

On speed, Flash streams near 300 output tokens per second with task times near 2.5 minutes at high reasoning and 48 seconds at low, while time to first token regressed to about 13 seconds with p95 observations near 22 seconds. Spark shows first-token latency near 2.7 to 3.0 seconds with throughput near 79 tokens per second on OpenRouter and a 38.5-second average time to first answer token on Artificial Analysis. The two speed profiles differ: Flash starts slow and streams fast, Spark starts fast with steadier output.

high-volume batch-friendly workloads favor Flash under promo pricing with caching and low thinking levels configured. Sustained agentic workloads at higher intelligence favor Spark standard on verified task cost, and Spark contributor undercuts everything where data terms allow. Annual budgets must model Flash's January price doubling explicitly.

Material unknowns to carry into procurement

All vendor-table figures above are provider runs unless attributed to Artificial Analysis, LMArena, or hands-on tests. Three presentation caveats deserve weight: Meta's scorecard places Spark 1.3 max beside Spark 1.2 xhigh, so generational jumps partly reflect reasoning-tier differences; the max figures describe a tier still in limited preview; and Terminal-Bench version selection flips conclusions, with Flash tied near the top on 2.1 and far behind on 4.0. Confirm Flash post-promo rates, Spark max availability and pricing, contributor data-use terms, and cutoff coverage for team domains before committing.

Choosing between them

Codebase engineering, long-context retrieval, and general knowledge work point to Muse Spark 1.3. Finance, legal, lab-science, and video-multimodal analysis at high volume point to Gemini 3.8 Flash under promotional pricing. Latency-sensitive chat with fast first tokens favors Spark, bulk streaming throughput favors Flash. Privacy-constrained teams use Spark standard or Flash standard; cost-maximal teams evaluate Spark contributor only after clearing data-use review.

Running both models inside a product with Ginger Labs

Affordable-model leadership is rotating monthly, so product teams gain from treating the model as a replaceable component behind a stable integration layer.

Ginger Labs ships an embedded agent that lives inside a customer's SaaS or web application, reasons over that product's schemas, records, and data, and carries out multi-step work for end users. Teams define the agent's capabilities and route between Spark 1.3, Gemini 3.8 Flash, and fallback models without rebuilding the product integration each time pricing or checkpoints change. The customer keeps ownership of its API, data model, permissions, domain rules, and definition of a correct result, including which actions each model may attempt and which data tiers each endpoint may receive.

For product capabilities that external clients should reach, Ginger Labs ships managed MCP infrastructure that exposes selected tools under customer-governed access, so Claude, Gemini, Spark-driven, and future clients connect without the team operating a separate MCP server per model family. A practical next step is a 20-minute demo scoped to one valuable workflow, run in a sandbox of the product, with both affordable models routed side by side on recorded traces.

Sources

• Google, Gemini 3.8 Flash and 3.8 Flash Cyber launch announcement, September 2, 2026, accessed September 2026
• Google DeepMind, Gemini 3.8 Flash model card, September 2026, accessed September 2026
• Meta AI Research, Muse Spark 1.3 launch announcement, September 2026, accessed September 2026
• Artificial Analysis, Muse Spark 1.3 evaluation, September 2, 2026, accessed September 2026
• Artificial Analysis, Gemini 3.8 Flash release intelligence, September 4, 2026, accessed September 2026
• Ars Technica, Gemini 3.8 Flash release analysis, September 2, 2026, accessed September 2026
• The Decoder, Gemini 3.8 Flash pricing and benchmark analysis, September 2, 2026, accessed September 2026
• Coursiv, Gemini 3.8 Flash benchmarks and pricing breakdown, September 2, 2026, accessed September 2026
• Codersera, Gemini 3.8 Flash specs and thinking-token cost analysis, September 3, 2026, accessed September 2026
• eesel AI, Muse Spark 1.3 benchmarks and pricing analysis, September 3, 2026, accessed September 2026
• GLB GPT, Muse Spark 1.3 review with hands-on coding tests, September 3, 2026, accessed September 2026
• BenchLM, Muse Spark 1.3 benchmark ledger, September 2026, accessed September 2026