diff --git a/.github/workflows/run-bot-aib-tournament.yaml b/.github/workflows/run-bot-aib-tournament.yaml index 92c4633e..077c4c6f 100644 --- a/.github/workflows/run-bot-aib-tournament.yaml +++ b/.github/workflows/run-bot-aib-tournament.yaml @@ -86,6 +86,64 @@ 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) + #################################### August 2026 new bots #################################### + + bot_muse_spark_1_2: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_MUSE_SPARK_1_2" + metac_name: "metac-muse-spark-1-2+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: ${{ secrets.ASKNEWS_SECRET }} + + bot_qwen_3_8_max: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_QWEN_3_8_MAX" + metac_name: "metac-qwen-3-8-max+asknews" + cache_key: asknews-cache-${{ 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: ${{ secrets.ASKNEWS_SECRET }} + + bot_inkling_small: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_INKLING_SMALL" + metac_name: "metac-inkling-small+asknews" + cache_key: asknews-cache-${{ 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: ${{ secrets.ASKNEWS_SECRET }} + + bot_deepseek_v4_flash: + needs: precache_asknews + uses: ./.github/workflows/run-bot-launcher.yaml + with: + bot_name: "METAC_DEEPSEEK_V4_FLASH" + metac_name: "metac-deepseek-v4-flash+asknews" + cache_key: asknews-cache-${{ 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 }} + #################################### July 2026 new bots #################################### bot_claude_opus_5: diff --git a/run_bots.py b/run_bots.py index 6507a7d9..a03a694c 100644 --- a/run_bots.py +++ b/run_bots.py @@ -619,6 +619,48 @@ def get_default_bot_dict() -> dict[str, RunBotConfig]: # NOSONAR } mode_base_bot_mapping = { + ############################ Bots started in August 2026 ############################ + "METAC_MUSE_SPARK_1_2": { + "estimated_cost_per_question": roughly_gpt_5_cost * 0.425, + "bot": create_bot( + llm=GeneralLlm( + model="openrouter/meta/muse-spark-1.2", + temperature=default_temperature, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_QWEN_3_8_MAX": { + "estimated_cost_per_question": roughly_gpt_5_cost * 0.6, + "bot": create_bot( + llm=GeneralLlm( + model="openrouter/qwen/qwen3.8-max", + temperature=default_temperature, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_INKLING_SMALL": { + "estimated_cost_per_question": roughly_gpt_5_cost * 0.12, + "bot": create_bot( + llm=GeneralLlm( + model="openrouter/thinkingmachines/inkling-small", + temperature=default_temperature, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, + "METAC_DEEPSEEK_V4_FLASH": { + "estimated_cost_per_question": roughly_deepseek_r1_cost * 0.072, + "bot": create_bot( + llm=GeneralLlm( + model="openrouter/deepseek/deepseek-v4-flash-0731", + temperature=default_temperature, + timeout=5 * 60, + ), + ), + "tournaments": TournConfig.aib_and_site, + }, ############################ Bots started in July 2026 ############################ "METAC_CLAUDE_OPUS_5": { "estimated_cost_per_question": roughly_opus_4_5_cost,