Arena.ai: Agent, Text, and WebDev rankings at High

Historical snapshot notice

The body preserves research collected on September 8, 2026. The editorial summary and conditions above were reviewed on September 20. References to “current” values, rankings, and dynamic metrics in these notes describe the original snapshot, not live results. Reopen the source before making a decision.

One-sentence takeaway

In Arena.ai's 2026-09-03 launch snapshot, Gemini 3.8 Flash (High) ranked 14th on Agent Arena (net improvement +5.94%), 7th on Text Arena (1,494 points), and 18th on Code Arena: WebDev (1,567 points). It was competitive on the cost-performance frontier for Agent tasks, but these results are specific to Arena's tasks and voting window and cannot be generalized into an overall ranking across all tasks.

Use cases

• Tasks suited for observation: Real-world long-horizon agents that call web, filesystem, and terminal tools; open-ended text-to-text tasks; and WebDev/frontend development workflows requiring multi-step reasoning and tool use.

• Tasks not suitable for direct inference: General knowledge, software engineering outside WebDev, production-system stability, latency, and end-to-end success rate; the Arena pages provide no unified conclusion on these tasks.

• Applicable model version: Gemini 3.8 Flash (High). Do not extend the conclusions to Gemini 3.8 Flash Cyber, other reasoning levels, or unlabeled snapshots.

• Applicable client, agent, or API: Arena's Agent Arena, Text Arena, and Code Arena: WebDev evaluation environments; these are not end-to-end scores from Gemini App, Claude Code, or a custom harness.

• Recommended parameters: The original posts did not disclose temperature, system prompt, tool schema, retry policy, or the complete sampling configuration; when retesting, these variables should be fixed rather than filled in as “default configuration.”

Test environment

• Agent Arena: The platform describes this as measuring how models orchestrate tools to complete real-world agent tasks, using web, filesystem, and terminal tools; core signals include Confirmed Success, Praise vs. Complaint, Steerability, Bash Recovery, and Tool Hallucination.

• Text Arena: The platform describes this as text-to-text tasks spanning open domains such as mathematics, coding, and creative writing; the X launch post reported 1,494 points and 7th place for Gemini 3.8 Flash (High).

• Code Arena: WebDev: The platform describes this as frontend web development tasks, including agentic coding workflows requiring multi-step reasoning and tool use; the X launch post reported 1,567 points and 18th place.

• Version and pricing: Arena labels the model Google · Proprietary; input/output pricing is $0.75 / $3.75 per million tokens.

Inputs/configuration

• The X posts did not disclose the Agent Arena task set, original prompts, number of model calls, random seeds, or complete harness configuration.

• The 2026-09-03 Agent post gave a median cost of $0.22 per task; the same post compared Gemini 3.8 Flash (High)'s +5.94% with Grok 4.5 at +6.17% / $0.39 and GLM 5.2 (Max) at +6.23% / $0.44.

• The prices are the model unit prices and platform cost estimates recorded on Arena's pages, not a promise covering every provider, cache-hit rate, or user bill.

Results data

X launch snapshot (2026-09-03)

Comparison of the five Agent signals (3.8 versus 3.7 in the X post):

• Praise vs. Complaint: +14.78% vs. -1.60%;

• Confirmed Success: +11.12% vs. +9.80%;

• Bash Recovery: +4.46% vs. +0.61%;

• Steerability: -1.43% vs. -5.38%;

• Tool Hallucination: the post said neither showed an issue.

Leaderboard refresh snapshot visible on the collection date

The leaderboards are dynamic pages, and the time windows shown on the collection date differed from those in the launch posts:

Therefore, the #14 / #7 / #18 figures in the X posts are launch-time snapshots, while the current-page #17 / #8 / #21 figures are later refresh results; they must not be combined as rankings from the same point in time. The rank movement itself also shows that the evaluation changes as sessions and votes are added.

Conclusions

The Arena launch snapshot supports three limited judgments:

1. On the platform's long-horizon Agent tasks, Gemini 3.8 Flash (High) showed a clear jump over Gemini 3.7 Flash (High), entering the Arena-labeled cost/performance frontier at a median task cost of $0.22.

2. Text Arena's 1,494 points placed it in the top ten at launch and above Claude Opus 5 (High) and Gemini 3.7 Flash (High) listed in that post; this indicates only the result for that Text Arena window.

3. WebDev's 1,567 points represent a usable upper-middle position, but not a leading one; on the refreshed page, Gemini 3.7 Flash was instead above it, so the “Agent improvement” should not be recast as “improvement on all coding tasks.”

Limitations

• Arena uses community votes and real-user task signals rather than a controlled laboratory benchmark; the X posts did not disclose a complete task set, pairwise randomization, whether blinding was strict, prompts, temperature, retries, or raw scoring logs. The “blind test” in the filename should not be interpreted as a verified double-blind experimental design.

• The three leaderboards use different sample units and time windows: Agent uses sessions, while Text/WebDev use votes; the page dates are September 6, September 3, and September 5, respectively.

• The launch rankings in the X posts had already drifted from the pages on the collection date, showing that rankings change with the sample; treat them as dated snapshots rather than permanent positions.

• Agent net improvement, signal percentages, and cost use Arena's platform definitions and do not equal the success rate or bill for a team using the same model, tools, and context.

• Some models in the Text/WebDev rows are labeled Preliminary, and the pages show confidence intervals or rank spreads; small differences between adjacent ranks should not be read as stable wins or losses.

• The results cover only three categories of Arena tasks—Agent, Text, and WebDev—and cannot be used to evaluate visual understanding, video, multilingual performance, safety, or all software-engineering work.

Reproduction steps

1. Save the access dates for the three original X posts and the three leaderboards, recording the page label date, total sessions/votes, model row, price, and confidence interval separately; do not replace the launch rankings with refreshed rankings.

2. For an independent retest, fix the Gemini 3.8 Flash (High) model snapshot, system prompt, tool schema, context limit, temperature, retry/stop conditions, and cost-calculation method.

3. Build separate task sets for Agent, Text, and WebDev; Agent tasks should cover at least web, filesystem, terminal, and error recovery, while WebDev tasks should have separate functional acceptance and visual regression checks.

4. Save the original input, tool trace, final artifact, human judgment, and token/cost data for every task; report success rates and confidence intervals rather than merely repeating Arena ranks.

5. Test Gemini 3.7 Flash (High) under the same conditions and limit the results to the corresponding task set; do not treat Arena's ranking as an overall model score for all tasks.

Raw evidence and data

• The @arena launch post explicitly stated that Gemini 3.8 Flash (High) had entered Agent Arena, Text Arena, and Code Arena: WebDev for the first time, and gave #14 / +5.94%, 1,494 / #7, and 1,567 / #18.

• The @arena Agent metrics post gave the five signal differences versus Gemini 3.7 Flash (High) and explained that Agent Arena uses web, filesystem, and terminal tools to measure real-world long-horizon tasks.

• The @arena Pareto post gave the $0.22-per-task figure and the cost/net-improvement comparison with Grok 4.5 and GLM 5.2 (Max).

• The Arena pages opened directly on the collection date provided later dynamic snapshots: 2,285,256 sessions / 59 models for Agent; 7,999,020 votes / 400 models for Text; and 650,961 votes / 126 models for WebDev, along with the current row data for Gemini 3.8 Flash (High).

Source excerpts or observations (for compliant short quotation only)

The X launch post summarized it as “+5.94% net improvement,” while the Pareto post said it had entered the cost/performance frontier of Agent Arena; these two statements support conclusions only for the corresponding Arena windows and cannot replace cross-task retesting.