Claude Opus 4.8 tops Appwrite Arena: the June 2026 leaderboard update_

Appwrite Arena is an open-source benchmark that measures how well AI models understand Appwrite. It scores each model on 191 questions spanning every Appwrite service, run twice: once with the relevant Appwrite Skill loaded into context, and once on the model's training knowledge alone. The gap between those two runs is what tells you how well a model already knows the platform. The June update adds four new frontier models, taking the board from 11 to 15, and one of them, Claude Opus 4.8, takes first place on the without-skills leaderboard.

Claude Opus 4.8 leads the without-skills leaderboard

On the without-skills board, where models answer from training knowledge alone with no Appwrite documentation in the prompt, Claude Opus 4.8 scores 97.4% overall and takes first place. It is the first model to clear 97% in that mode, and the first to rank above Claude Opus 4.7.

For almost every model on the board, adding Appwrite documentation to the prompt raises the score, because the documentation closes a knowledge gap. Claude Opus 4.8 is the first model where that does not hold: it scores higher without skills (97.4%) than with them (97.1%). The model already knows Appwrite well enough from training that adding documentation to the prompt does not improve its accuracy.

The same pattern appears in cost. At $5 per million input tokens, including the skills documentation in every prompt raises the with-skills run to $6.86, more than four times the $1.56 without-skills run. For Claude Opus 4.8, skills add cost and slightly lower the score, making it the first model on the board better run without them.

New models added in June 2026

Claude Opus 4.8 is not the only addition. Three other frontier models also joined since May, each with a different balance of speed and cost.

Grok Build 0.1

• Ranks fourth with skills at 96.7%, running at 138 tok/s, far above Kimi K2.6's 17 tok/s.
• Its free-form score gains 7.5 points with skills, from 83.7% to 91.2%.
• Priced at $1.00 / $2.00 per million tokens, or $2.28 per with-skills run.

Gemini 3.5 Flash

• Ranks seventh with skills at 96.2% and runs at 118 tok/s.
• Depends most on documentation of the new models: overall falls from 96.2% with skills to 90.7% without, and free-form moves 14.4 points, from 77.5% to 91.9%.
• At $9.00 per million output tokens, a with-skills run costs $3.78, among the higher figures on the board.

MiniMax M3

• Offers the strongest cost-to-score ratio: $0.49 per with-skills run (95.7%) and $0.09 without skills (91.0%).
• Its 95.2% free-form is the highest of the four new models.
• A clear improvement over MiniMax M2.7: 93.2% to 95.7% with skills, and 85.2% to 91.0% without.
• Its $0.30 / $1.20 per-million pricing reflects a 50% discount on OpenRouter running until June 7, 2026, so the cost figures above will rise once it ends.

Without-skills leaderboard rankings

Adding Claude Opus 4.8 reorders the top of the without-skills rankings, where the spread between models is widest.

The top of the without-skills board now reads:

Two Anthropic models now hold the top two positions without any documentation, with GPT 5.5 close behind. The free-form column shows the expected pattern: the models that drop the most without skills are those that rely on documentation to answer open-ended questions, and the gap between multiple-choice and free-form widens further down the table.

With-skills leaderboard rankings

With Appwrite documentation in the prompt, the board compresses toward the top. Ten of the fifteen models score 95.7% or higher, and the top six sit within 1.4 points of each other.

• GPT 5.5 holds first place at 97.7%, the only model above 97.5% with skills, on the strength of a board-leading 98.2% on multiple-choice.
• The two Anthropic models trade places from the without-skills board. With skills, Claude Opus 4.7 ranks #2 and Claude Opus 4.8 ranks #3, both at 97.1% with identical multiple-choice scores (97.6%) and 186 of 191 correct. Without skills the order is reversed, with Opus 4.8 at 97.4% ahead of Opus 4.7 at 96.2%. Documentation lifts Opus 4.7 by 0.9 points (96.2% to 97.1%) but does not help Opus 4.8 (97.4% to 97.1%), so the two converge once the docs are in the prompt.
• The field stays tight below the top. Grok Build 0.1 (96.7%), Qwen 3.6 Plus (96.5%), and Kimi K2.6 (96.3%) are separated by fractions of a point, so cost and speed, rather than accuracy, decide between them.

Resources

The Arena UI lets you filter by category, switch between with and without skills, sort by any column, and click through to a per-model breakdown with per-question reasoning and tool call counts. The repo is open source, so you can re-run the benchmark locally against your own OpenRouter key.