Menu

TAO Synergies · Nasdaq: TAOX Powered by Bittensor

xagram

xagram is a Solution Intelligence system. It picks the best AI for every task, checks its facts, and actually builds and runs things. In our October 2026 runs against seven leading models, it placed #1 on AutomationBench (ahead of Claude Opus 5.5, ChatGPT Astra, and Kimi K3.1), tied for #1 on AA-Omniscience with Claude Opus 5.5, and MATH-500, and scored ahead of Kimi K3.1 on TerminalBench and was #1 on SimpleQA. The API is open: get a key and start in two lines.

Updated

Web BuildsSee every xagram Gallery site next to Claude Fable 5.1 →Same asks, side by side, with build times, costs and live links.

Pricing

Chat: free. New accounts get $5 of credits for Build and Video Studio. Build and API: $0.40 / $1.95 per million input / output tokens.

Price per million tokensInputOutput
xagram$0.40$1.95
Claude Fable 5.1$10.00$50.00
GPT-6 Astra$10.00$50.00
Claude Opus 5.5$4.00$20.00
Kimi K3.1 Pro$2.00$10.00

xagram: $0.40 / M input · $0.12 / M cached input · $1.95 / M output. OpenAI-compatible API.

Other models: each provider's published standard API price per million tokens, read Oct 26, 2026, before cached-input, batch or other discounts: Anthropic for Claude Fable 5.1 and Claude Opus 5.5 (anthropic.com/pricing), OpenAI for GPT-6 Astra (platform.openai.com/docs/pricing), Moonshot AI for Kimi K3.1 Pro (platform.kimi.ai/docs/pricing/chat).

November 2026: xagram

BenchmarkResults
TERMINAL-BENCH 2.1Full 87-task set. Published scores are mostly self-reported (BenchLM, Nov 5, 2026). xagram: all 87 tasks, mean of 3 attempts.Cognition SWE-2 92.8%GPT-5.6 Sol 91.9%xagram 90.8% (mean of 3 attempts on all 87 tasks: 79, 80, 78)
AUTOMATIONBENCH, held-outDifferent task sets, so not directly comparable. xagram: a sample of 40 held-out tasks, 3 runs each (120 trials), strict grading.xagram 63.3% (76 of 120, 40-task sample, 3 runs each)MiniMax M3.1 Flash 62.8% (company-reported)Gemini 4 Carbon 53.9% (Zapier independent, private set)
xagram runs on Nov 12, 2026. Terminal-Bench 2.1: all 87 tasks, mean of 3 attempts. AutomationBench held-out is a sample (40 tasks, 3 runs each), not a full benchmark run. Raw result files (md5, first 8 hex): AutomationBench held-out 40 823d6747, ccb21667, b6ce2552.

Earlier results (October 2026)

BenchmarkxagramBest rivalAhead ofWhere xagram stands
AA-OMNISCIENCE600 public questions; knowledge index, −100 to 100. Our run (Oct 26), not Artificial Analysis’s own Intelligence Index62.7Oct 26 version62.0Claude Opus 5.5Next: 58.3 GPT-6 Astra, GPT-6.1 Sol58.3GPT-6 Astra, GPT-6.1 SolTied #1 of 8
AUTOMATIONBENCHfirst 100 public AutomationBench 1.0.6 tasks, one run each75.0%agent setup69.0%Kimi K3.1 ProZapier leaderboard (private held-out set, v1.0.6): Claude Opus 5.5 42.5%, GPT-6 Astra 41.4%; ours is the first 100 public tasks69.0%Kimi K3.1 ProAhead of Kimi K3.1
TERMINALBENCHTerminalBench 4.0; the tasks both models completed, one run, 90-minute limit per task64.3%xagram Max42.9%Kimi K3.1 Pro (our run)Published, different harness and full task set: Claude Opus 5.5 64.9%, GPT-6 Astra 58.2%42.9%Kimi K3.1 ProAhead of Kimi K3.1
SIMPLEQA500 factual questions; xagram with web search, the rivals without84.4%Oct 26 version72.2%GPT-6 Astra, GPT-6.1 SolNext: 67.0% Claude Opus 5.572.2%GPT-6 Astra, GPT-6.1 Sol#1 of 8
MATH-500500 competition math problems, strict grading98.6%Oct 26 version99.4%Claude Opus 5.5Next: 98.6% Gemini 4 Argon94.8%GPT-6 AstraTied #1 of 9
GPQA DIAMOND198 graduate-level science questions, one attempt each95.5%Oct 26 version96.5%Gemini 4 ArgonNext: 95.5% GPT-6 Astra, GPT-6.1 Sol93.4%Kimi K3.12nd of 9
IFEVAL200 prompts, strict prompt-level scoring95.8%average of 2 runs97.0%GPT-6 AstraNext: 95.5% Kimi K3.1Published: no IFEval score found for Claude Fable 5.1 or GPT-6.1 Sol93.0%Claude Opus 5.5Tied 2nd of 8
Our runs through each model's API on the same questions with the same grader, Oct 2026; a tie is within 1 point. Rows marked "published" use another lab's numbers on a different task set or harness and are not directly comparable.
Methods and raw files

How we measured: xagram numbers are our runs through the public xagram API with default settings on Oct 26, 2026 (IFEval: the average of two runs on the same 200 prompts, Oct 25 and Oct 26; SimpleQA rows span several API versions that day), except AutomationBench and TerminalBench, which were measured with xagram's agent setup in our harness. Rival numbers are our runs through each model's API on the same items with the same grader, unless marked "published": Kimi K3.1, GPT-6 Astra, GPT-6.1 Sol, Claude Sonnet 5.5, Claude Opus 5.5, DeepSeek 1017 and Mistral Large 4, plus Gemini 4 Argon on GPQA Diamond and MATH-500. A row is a tie when the two scores are within 1 point. "Ahead of" shows the highest-scoring of GPT-6 Astra, GPT-6.1 Sol, Claude Opus 5.5, Gemini 4 Argon and Kimi K3.1 that xagram beats by more than 1 point on the same items (— when none; Claude Fable 5.1 was not run by us). AA-Omniscience: graded with AA's grader (GPT-5.6 Luna, AA's template) on Oct 26, before Artificial Analysis released Intelligence Index v5 on Oct 28; AA's own published index uses its full question set (v5 is now mostly private datasets) and is not comparable. SimpleQA: xagram used web search, the rivals did not. IFEval: Kimi K3.1 (95.5%), GPT-6.1 Sol (95.0%) and DeepSeek 1017 (95.0%) are within 1 point of xagram; we found no published IFEval score for Claude Fable 5.1 or GPT-6.1 Sol (BenchLM's IFEval leaderboard and model pages, read Oct 26, 2026). TerminalBench 4.0: both models ran the same 20 pre-selected tasks with a 90-minute limit per task (TerminalBench's standard limit is 8 hours); Kimi K3.1 Pro, in the standard Mini-SWE-agent harness, completed 14 before the run was stopped early, and the row compares the two on those 14; on all 20 tasks xagram Max scored 60.0%; 75% of xagram Max's passes reproduced on a rerun, and a 3-attempt run is in progress. The published TerminalBench scores are the Terminal-Bench 4.0 leaderboard snapshot of Oct 25, 2026 (tbench.ai, mirrored by BenchLM): Claude Opus 5.5 64.85% in Claude Code and GPT-6 Astra 58.18% in Codex, at maximum reasoning, on the full task set with standard limits. Raw results (path, md5): AA-Omniscience bench_finish_2026-10-26/AA_live_1026_2005_600_rows.jsonl da7946f557e62d2156fb52e00dbe32c2 (field grade_luna), Claude Opus 5.5 leaderboard_2026-10-26/runs/opus/omni_full/rows.jsonl ff752c34740398d96512e33e745b0ff1; SimpleQA bench_finish_2026-10-26/simpleqa_live_1026_2005/rows.jsonl a51700993c410b8939e5d652ef5d83a7; MATH-500 math500_live_1026_2005_math 6a610eca86c2e6eb23a0333d06ef7494, Gemini 4 Argon 5e593dab61708e90c59df0ad6a948820; GPQA Diamond suites_tao1_vs_k31_2026-10-25/exam3_gpqa_v12/rows.jsonl 595eaa94cbd752e6d0646f61df36ba98, Gemini 4 Argon runs/argon/gpqa_full d042df94fc1293e08e3f082357593b0b; IFEval ifeval200_live_1026_2005 a7356ca115eb5edf80f292c81e9c2a2c and suites_tao1_vs_k31_2026-10-25/full_ifeval200_tao1/rows.jsonl bf6377e0bc8cf727faac03936e18e3ee; all rivals leaderboard_2026-10-26/FINAL_TABLE.md a4fc511550472a235ac6b38c58fd8792; AutomationBench run 1 ab6d708a37c9b2b660d376852e0d934f (75), run 2 bec16a6c92c6b9a869dacab1650b88a6 (74), Kimi K3.1 Pro 63e3dbd7dd02742518fb500466bd5d3e; TerminalBench handover/TB_K31_TBRULES_vs_MINISWE_v2_20of20_2026-10-26.json e3995f37c8a7175ecd15474993eef88e.

Web Builds: xagram Gallery sites next to Claude Fable 5.1 · Earlier results (Oct 2026)