diff --git a/.github/workflows/run-bot-aib-tournament.yaml b/.github/workflows/run-bot-aib-tournament.yaml index fb849ce7..3c712492 100644 --- a/.github/workflows/run-bot-aib-tournament.yaml +++ b/.github/workflows/run-bot-aib-tournament.yaml @@ -86,28 +86,145 @@ jobs: # NOTE: don't remove any of the open source models, since these are the best option for a long term baseline (other models get deprecated) - bot_grok_4_1_high: # TODO: Not yet released via API as of Dec 21st, 2025 + #################################### April 2026 new bots #################################### + + bot_claude_sonnet_4_6: needs: precache_asknews uses: ./.github/workflows/run-bot-launcher.yaml with: - bot_name: "METAC_GROK_4_1_HIGH" + bot_name: "METAC_CLAUDE_SONNET_4_6" + metac_name: "metac-claude-sonnet-4-6+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: - INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GROK_4_1_HIGH }} - INPUT_XAI_API_KEY: ${{ secrets.XAI_API_KEY }} + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} + INPUT_ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }} + INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} + INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} + INPUT_OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }} + + bot_qwen_3_5: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_QWEN_3_5" + metac_name: "metac-qwen-3-5+asknews" + cache_key: asknews-cache-${{ github.run_id }} + secrets: + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} INPUT_OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }} INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} - bot_grok_4_1: # TODO: Not yet released via API as of Dec 21st, 2025 + bot_gemini_3_1_pro: needs: precache_asknews uses: ./.github/workflows/run-bot-launcher.yaml with: - bot_name: "METAC_GROK_4_1" + bot_name: "METAC_GEMINI_3_1_PRO" + metac_name: "metac-gemini-3-1-pro+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: - INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GROK_4_1 }} - INPUT_XAI_API_KEY: ${{ secrets.XAI_API_KEY }} + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} + INPUT_OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }} + INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} + INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} + + bot_gpt_5_5: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_GPT_5_5" + metac_name: "metac-gpt-5-5+asknews" + cache_key: asknews-cache-${{ github.run_id }} + secrets: + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} + INPUT_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }} + INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} + INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} + + bot_gpt_5_4: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_GPT_5_4" + metac_name: "metac-gpt-5-4+asknews" + cache_key: asknews-cache-${{ github.run_id }} + secrets: + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} + INPUT_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }} + INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} + INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} + + bot_gpt_5_4_mini: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_GPT_5_4_MINI" + metac_name: "metac-gpt-5-4-mini+asknews" + cache_key: asknews-cache-${{ github.run_id }} + secrets: + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} + INPUT_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }} + INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} + INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} + + bot_gpt_5_4_nano: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_GPT_5_4_NANO" + metac_name: "metac-gpt-5-4-nano+asknews" + cache_key: asknews-cache-${{ github.run_id }} + secrets: + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} + INPUT_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }} + INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} + INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} + + bot_gpt_5_3: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_GPT_5_3" + metac_name: "metac-gpt-5-3+asknews" + cache_key: asknews-cache-${{ github.run_id }} + secrets: + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} + INPUT_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }} + INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} + INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} + + bot_gpt_5_3_instant: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_GPT_5_3_INSTANT" + metac_name: "metac-gpt-5-3-instant+asknews" + cache_key: asknews-cache-${{ github.run_id }} + secrets: + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} + INPUT_OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }} + INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} + INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} + + bot_minimax_m2_7: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_MINIMAX_M2_7" + metac_name: "metac-minimax-m2-7+asknews" + cache_key: asknews-cache-${{ github.run_id }} + secrets: + INPUT_METACULUS_TOKENS: ${{ secrets.METACULUS_TOKENS }} + INPUT_METACULUS_API_BASE_URL: ${{ secrets.METACULUS_API_BASE_URL }} INPUT_OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }} INPUT_ASKNEWS_CLIENT_ID: ${{ secrets.ASKNEWS_CLIENT_ID }} INPUT_ASKNEWS_SECRET: ${{ secrets.ASKNEWS_SECRET }} @@ -119,6 +236,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_CLAUDE_OPUS_4_6_HIGH_32K" + metac_name: "metac-claude-opus-4-6-high-32k+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_CLAUDE_OPUS_4_6_HIGH_32K }} @@ -147,6 +265,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_CLAUDE_HAIKU_4_5" + metac_name: "metac-claude-haiku-4-5+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_CLAUDE_HAIKU_4_5 }} @@ -161,6 +280,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_QWEN3_MAX_THINKING" + metac_name: "metac-qwen3-max-thinking+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_QWEN3_MAX_THINKING }} @@ -187,6 +307,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_KIMI_K2_5_HIGH" + metac_name: "metac-kimi-k2-5+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_KIMI_K2_5_HIGH }} @@ -200,6 +321,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GLM_5" + metac_name: "metac-glm-5+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GLM_5 }} @@ -215,6 +337,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_CLAUDE_OPUS_4_5_HIGH_32K" + metac_name: "metac-claude-opus-4-5-high-32k+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_CLAUDE_OPUS_4_5_HIGH_32K }} @@ -229,6 +352,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_CLAUDE_OPUS_4_5" + metac_name: "metac-claude-opus-4-5+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_CLAUDE_OPUS_4_5 }} @@ -243,6 +367,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_5_2_HIGH" + metac_name: "metac-gpt-5-2-high+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_5_2_HIGH }} @@ -257,6 +382,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_5_2" + metac_name: "metac-gpt-5-2+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_5_2 }} @@ -284,6 +410,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GEMINI_3_FLASH" + metac_name: "metac-gemini-3-flash+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GEMINI_3_FLASH }} @@ -298,6 +425,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GLM_4_6" + metac_name: "metac-glm-4-6+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GLM_4_6 }} @@ -325,6 +453,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_KIMI_K2_HIGH" + metac_name: "metac-kimi-k2-high+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_KIMI_K2_HIGH }} @@ -338,6 +467,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_5_1_HIGH" + metac_name: "metac-gpt-5-1-high+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_5_1_HIGH }} @@ -352,6 +482,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_5_1" + metac_name: "metac-gpt-5-1+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_5_1 }} @@ -393,6 +524,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GROK_4_1_FAST_HIGH" + metac_name: "metac-grok-4-1-fast-high+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GROK_4_1_FAST_HIGH }} @@ -407,6 +539,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GROK_4_1_FAST" + metac_name: "metac-grok-4-1-fast+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GROK_4_1_FAST }} @@ -447,6 +580,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_CLAUDE_4_5_SONNET_HIGH" + metac_name: "metac-claude-4-5-sonnet-high-32k+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_CLAUDE_4_5_SONNET_HIGH }} @@ -461,6 +595,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_CLAUDE_4_5_SONNET" + metac_name: "metac-claude-4-5-sonnet+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_CLAUDE_4_5_SONNET }} @@ -475,6 +610,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_QWEN_3_MAX" + metac_name: "metac-qwen-3-max+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_QWEN_3_MAX }} @@ -488,6 +624,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_3_2_REASONING" + metac_name: "metac-deepseek-3-2-reasoning+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_3_2_REASONING }} @@ -513,6 +650,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GROK_4_FAST_HIGH" + metac_name: "metac-grok-4-fast-high+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GROK_4_FAST_HIGH }} @@ -569,6 +707,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_5_MINI" + metac_name: "metac-gpt-5-mini+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_5_MINI }} @@ -583,6 +722,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_5_NANO" + metac_name: "metac-gpt-5-nano+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_5_NANO }} @@ -636,6 +776,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GROK_4" + metac_name: "metac-grok-4+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GROK_4 }} @@ -650,6 +791,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_KIMI_K2" + metac_name: "metac-kimi-k2+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_KIMI_K2 }} @@ -676,6 +818,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_VARIANCE_TEST" + metac_name: "metac-deepseek-r1+asknews(variance-test)" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_VARIANCE_TEST }} @@ -689,6 +832,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_OSS_120B" + metac_name: "metac-gpt-oss-120b+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_OSS_120B }} @@ -702,6 +846,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_ZAI_GLM_4_5" + metac_name: "metac-zai-glm-4-5+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_ZAI_GLM_4_5 }} @@ -715,6 +860,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_V3_1_REASONING" + metac_name: "metac-deepseek-v3-1-reasoning+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_V3_1_REASONING }} @@ -728,6 +874,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_V3_1" + metac_name: "metac-deepseek-v3-1+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_V3_1 }} @@ -741,6 +888,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_V3_1_VARIANCE_TEST_1" + metac_name: "metac-deepseek-v3-1+asknews(variance-test-1)" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_V3_1_VARIANCE_TEST_1 }} @@ -754,6 +902,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_V3_1_VARIANCE_TEST_2" + metac_name: "metac-deepseek-v3-1+asknews(variance-test-2)" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_V3_1_VARIANCE_TEST_2 }} @@ -812,6 +961,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GEMINI_2_5_PRO_GROUNDING" + metac_name: "metac-gemini-2-5-pro-grounding[research-only]" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GEMINI_2_5_PRO_GROUNDING }} @@ -824,6 +974,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_ASKNEWS_DEEPNEWS" + metac_name: "metac-asknews-deepnews[research-only]" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_ASKNEWS_DEEPNEWS }} @@ -870,6 +1021,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_EXA_ONLINE_RESEARCH_ONLY" + metac_name: "metac-deepseek-r1-exa-online[research-only-bot]" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_EXA_ONLINE_RESEARCH_ONLY }} @@ -882,6 +1034,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_PLUS_EXA_ONLINE" + metac_name: "metac-deepseek-r1+deepseek-r1-exa-online" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_PLUS_EXA_ONLINE }} @@ -927,6 +1080,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_NO_RESEARCH" + metac_name: "metac-deepseek-r1+no-research" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_NO_RESEARCH }} @@ -939,6 +1093,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_4_1_OPTIMIZED_PROMPT" + metac_name: "metac-gpt-4-1+asknews[optimized-prompt]" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_4_1_OPTIMIZED_PROMPT }} @@ -953,6 +1108,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_4_1_NANO_OPTIMIZED_PROMPT" + metac_name: "metac-gpt-4-1-nano+asknews[optimized-prompt]" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_4_1_NANO_OPTIMIZED_PROMPT }} @@ -1038,6 +1194,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_SONAR_PRO" + metac_name: "metac-deepseek-r1+sonar-pro" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_SONAR_PRO }} @@ -1050,6 +1207,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_SONAR" + metac_name: "metac-deepseek-r1+sonar" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_SONAR }} @@ -1073,6 +1231,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_SONAR_REASONING_PRO" + metac_name: "metac-deepseek-r1+sonar-reasoning-pro" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_SONAR_REASONING_PRO }} @@ -1108,6 +1267,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_GPT_4O_SEARCH_PREVIEW" + metac_name: "metac-deepseek-r1+gpt-4o-search-preview" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_GPT_4O_SEARCH_PREVIEW }} @@ -1120,6 +1280,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_GEMINI_2_5_PRO_GROUNDING" + metac_name: "metac-deepseek-r1+gemini-2-5-pro-grounding" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_GEMINI_2_5_PRO_GROUNDING }} @@ -1143,6 +1304,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_ASK_EXA_PRO" + metac_name: "metac-deepseek-r1+exa-answer" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_ASK_EXA_PRO }} @@ -1181,6 +1343,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_O3_TOKEN" + metac_name: "metac-o3+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_O3_TOKEN }} @@ -1195,6 +1358,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_O4_MINI_HIGH_TOKEN" + metac_name: "metac-o4-mini-high+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_O4_MINI_HIGH_TOKEN }} @@ -1209,6 +1373,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_O4_MINI_TOKEN" + metac_name: "metac-o4-mini+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_O4_MINI_TOKEN }} @@ -1223,6 +1388,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_4_1_TOKEN" + metac_name: "metac-gpt-4-1+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_4_1_TOKEN }} @@ -1237,6 +1403,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_4_1_MINI_TOKEN" + metac_name: "metac-gpt-4-1-mini+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_4_1_MINI_TOKEN }} @@ -1251,6 +1418,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_4_1_NANO_TOKEN" + metac_name: "metac-gpt-4-1-nano+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_4_1_NANO_TOKEN }} @@ -1265,6 +1433,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GEMINI_2_5_FLASH_PREVIEW_TOKEN" + metac_name: "metac-gemini-2-5-flash+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GEMINI_2_5_FLASH_PREVIEW_TOKEN }} @@ -1344,6 +1513,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_4O_TOKEN" + metac_name: "metac-gpt-4o+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_4O_TOKEN }} @@ -1358,6 +1528,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_4O_MINI_TOKEN" + metac_name: "metac-gpt-4o-mini+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_4O_MINI_TOKEN }} @@ -1372,6 +1543,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GPT_3_5_TURBO_TOKEN" + metac_name: "metac-gpt-3-5-turbo+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GPT_3_5_TURBO_TOKEN }} @@ -1464,6 +1636,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_LLAMA_4_MAVERICK_17B_TOKEN" + metac_name: "metac-llama-4-maverick-17b+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_LLAMA_4_MAVERICK_17B_TOKEN }} @@ -1477,6 +1650,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_QWEN_2_5_MAX_TOKEN" + metac_name: "metac-qwen-2-5-max+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_QWEN_2_5_MAX_TOKEN }} @@ -1490,6 +1664,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_R1_TOKEN" + metac_name: "metac-deepseek-r1+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_R1_TOKEN }} @@ -1503,6 +1678,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_DEEPSEEK_V3_TOKEN" + metac_name: "metac-deepseek-v3+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_DEEPSEEK_V3_TOKEN }} @@ -1516,6 +1692,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GROK_3_LATEST_TOKEN" + metac_name: "metac-grok-3+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GROK_3_LATEST_TOKEN }} @@ -1530,6 +1707,7 @@ jobs: uses: ./.github/workflows/run-bot-launcher.yaml with: bot_name: "METAC_GROK_3_MINI_LATEST_HIGH_TOKEN" + metac_name: "metac-grok-3-mini-high+asknews" cache_key: asknews-cache-${{ github.run_id }} secrets: INPUT_METACULUS_TOKEN: ${{ secrets.METAC_GROK_3_MINI_LATEST_HIGH_TOKEN }} diff --git a/run_bots.py b/run_bots.py index 35ccc195..5073d6e5 100644 --- a/run_bots.py +++ b/run_bots.py @@ -562,6 +562,92 @@ def get_default_bot_dict() -> dict[str, RunBotConfig]: # NOSONAR mode_base_bot_mapping = { # "METAC_GROK_4_1_HIGH": {} # TODO: Add these bots to github workflow. Its not yet released via API as of Dec 21st, 2025 # "METAC_GROK_4_1": {} + ############################ Bots started in April 2026 ############################ + "METAC_CLAUDE_SONNET_4_6": { + "estimated_cost_per_question": roughly_sonnet_4_cost, + "bot": create_bot( + llm=GeneralLlm( + model="anthropic/claude-sonnet-4-6", + temperature=default_temperature, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_QWEN_3_5": { + "estimated_cost_per_question": roughly_sonnet_3_5_cost / 2, + "bot": create_bot( + GeneralLlm( + model="openrouter/qwen/qwen3.5-397b-a17b", + temperature=default_temperature, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_GEMINI_3_1_PRO": { + "estimated_cost_per_question": roughly_gemini_2_5_pro_preview_cost, + "bot": create_bot( + GeneralLlm( + model="openrouter/google/gemini-3.1-pro-preview", + temperature=default_temperature, + timeout=gemini_default_timeout, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_GPT_5_5": { + "estimated_cost_per_question": roughly_gpt_5_high_cost, + "bot": create_bot( + llm=GeneralLlm( + model="openai/gpt-5.5", + temperature=None, + timeout=gpt_5_timeout, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_GPT_5_4": { + "estimated_cost_per_question": roughly_gpt_5_cost, + "bot": create_bot( + llm=GeneralLlm( + model="openai/gpt-5.4", + temperature=default_temperature, + timeout=gpt_5_timeout, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_GPT_5_4_MINI": { + "estimated_cost_per_question": roughly_gpt_4o_mini_cost, + "bot": create_bot( + llm=GeneralLlm( + model="openai/gpt-5.4-mini", + temperature=default_temperature, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_GPT_5_4_NANO": { + "estimated_cost_per_question": roughly_gpt_4o_mini_cost / 2, + "bot": create_bot( + llm=GeneralLlm( + model="openai/gpt-5.4-nano", + temperature=default_temperature, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + # "METAC_GPT_5_3": {} -> Not yet available in OpenAI API as of April 2026 + # "METAC_GPT_5_3_INSTANT": {} -> Not yet available in OpenAI API as of April 2026 + "METAC_MINIMAX_M2_7": { + "estimated_cost_per_question": roughly_deepseek_r1_cost, + "bot": create_bot( + GeneralLlm( + model="openrouter/minimax/minimax-m2.7", + temperature=default_temperature, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, ############################ Bots started in February 2026 ############################ "METAC_CLAUDE_OPUS_4_6_HIGH_32K": { "estimated_cost_per_question": roughly_opus_4_5_cost * 1.3,