DeepSeek and Meta Add High-Compute Variants to Agent Arena Within 48 Hours
DeepSeek’s V4-Pro-High entered Arena’s Agent Arena and Text Arena leaderboards on August 19. Meta’s Muse Spark 1.2 xHigh followed on August 20. The two additions are the fourth and fifth extended-reasoning variants to reach Agent Arena since June, when GPT-5.5 xHigh debuted at rank two.
DeepSeek released both V4-Pro-High and V4-Flash-High simultaneously via API on August 13. V4-Flash-High extends the flash tier — the 79.0% SWE-bench model priced at $0.14/M — with higher-compute inference. V4-Pro-High builds on V4-Pro, which holds 96.4% on SWE-bench Verified, the second-highest score on record behind Claude Opus 5’s 97.0%.
Meta’s Muse Spark 1.2 launched August 6 with async subagents and a replay-safe runtime for multi-step coding workflows. The xHigh variant adds extended reasoning on top of the base model, using the same naming convention as GPT-5.5 Thinking xHigh Effort — which runs at 80.2 overall on LiveBench and appears on Agent Arena as one of OpenAI’s primary benchmark-competitive products.
A Full Set of Effort Tiers
The August additions complete what is now a full roster of high-compute variants from the top labs in Agent Arena:
| Agent Arena Entry | Model | Lab |
|---|---|---|
| June 14, 2026 | GPT-5.5 xHigh | OpenAI |
| July 14, 2026 | GPT-5.6 Sol xHigh | OpenAI |
| August 18, 2026 | Inkling-Small | Thinking Machines |
| August 19, 2026 | DeepSeek V4-Pro-High | DeepSeek |
| August 20, 2026 | Muse Spark 1.2 xHigh | Meta |
Anthropic’s Fable 5 Max Effort and Grok 4.x thinking modes are already established in the leaderboard. The pattern is now consistent: labs submit extended-compute variants as their Agent Arena entries while base models serve as the production API tier.
What the Variants Cost
The extended-reasoning Agent Arena submissions are not cheap. GPT-5.5 xHigh launched at $3.00/M on input. Claude Fable 5 Max Effort costs $1.439 per successful task on LiveBench. DeepSeek V4-Pro-High pricing was not disclosed in the August 13 API update, but DeepSeek’s August 16 peak/off-peak billing change raised V4 Pro rates by up to 12x at peak hours.
Labs are spending margin to hold Agent Arena benchmark positions. Whether that translates to production value depends on what human voters in Agent Arena sessions are actually measuring: Fable 5 Max Effort leads LiveBench at 83.0 overall, but ranks third in Agentic Coding at 62.2% behind Claude Opus 5 Thinking (65.2%) and Smaug-Agentic (64.6%). Higher compute does not uniformly improve the task categories that matter most to developers.
DeepSeek V4-Pro-High Specifics
V4-Pro-High adds extended reasoning to a base model that already sits at near-maximum SWE-bench performance. The practical question for Agent Arena is whether the additional compute improves multi-step task planning and error recovery — the criteria human voters use — rather than raw code accuracy, where V4-Pro is already at ceiling.
V4-Flash-High’s Agent Arena entry has not yet been confirmed. The ticker registry shows both the pro-high and flash-high variants as distinct API entries as of August 13, but only V4-Pro-High appeared in the August 19 Arena changelog.
Arena ratings for DeepSeek V4-Pro-High and Muse Spark 1.2 xHigh were not available at publication.