Compare model context windows, published API prices, and workloads. Shortlist up to three models and review them side by side.
ManasFlow catalog snapshot · October 2026 · Verify provider terms before use
Adjust your token volume to compare standard API costs. Prices and catalog entries are an October 2026 snapshot.
Monthly estimate = input millions × input price + output millions × output price. Excludes caching, batch discounts, infrastructure, and taxes.
Post-trained version of Kimi K3 (2.8T total / 104B active MoE). Always-on thinking with reasoning_effort low/high/max (default max). Text, image and video input. Moonshot says full weights ship by Nov 4 2026; until then it is API-only, so openSource is false. Cyber-related training data is removed from the open-weight release.
Released Oct 7 2026. Built for high-volume, latency-sensitive work such as classification, routing, extraction and subagent tasks; fastest comparative latency in the current Claude lineup. Unlike Haiku 4.5 (200K context, extended thinking), it has a 1M context window, 128K max output and adaptive thinking with the effort parameter (default effort medium). It uses the newer tokenizer from Claude Opus 4.7 onward, so the same text counts as about 30% more tokens than on Haiku 4.5. Thinking blocks work only in the account that produced them or a linked account. Text and image input, up to 300K output on the Batch API with the output-300k-2026-03-24 beta header. API identifier claude-haiku-5-5 (anthropic.claude-haiku-5-5 on Bedrock), reliable knowledge cutoff June 2026, retirement not sooner than Oct 7 2027. Source: https://platform.claude.com/docs/en/models/haiku-5-5/overview (checked 2026-10-23).
Released Sep 29 2026 for complex coding, computer use and professional work at a lower cost than GPT-6 Astra; OpenAI's models page now lists it in the Sol slot of the GPT-6 family, at the same $2/$10 price as GPT-6 Sol. Adds Multi-agent in beta (the model can delegate work to subagents in a Responses API request) and, since Oct 8 2026, an Ultrafast mode (service_tier "ultrafast"). Text and image input, 922K max input tokens, reasoning effort low to max (no none or minimal). Use the Responses API for tool calling; Chat Completions works without tools. API identifier gpt-6.1-sol, knowledge cutoff Apr 30 2026. Sources: https://developers.openai.com/api/docs/models/gpt-6.1-sol and the OpenAI API changelog (checked 2026-10-23).
Released Sep 28 2026. Anthropic describes it as the best combination of speed and intelligence in the current lineup (comparative latency: fast). Adaptive thinking is on by default (default effort high); the lowest setting, between_tools, turns off up-front thinking. Five breaking changes affect code written for Sonnet 5: up-front thinking is turned off with between_tools, forced tool use returns an error, thinking blocks are tied to the model and conversation, computer_20251124 is not accepted on the Claude API and Google-managed Google Cloud, and the advisor tool rejects Opus 4.8, Opus 4.7, Sonnet 5 and Haiku 5.5 as advisors. Setting temperature, top_p or top_k returns a 400 error. Text and image input, up to 300K output on the Batch API with the output-300k-2026-03-24 beta header. API identifier claude-sonnet-5-5 (anthropic.claude-sonnet-5-5 on Bedrock), reliable knowledge cutoff June 2026, retirement not sooner than Sep 28 2027. Source: https://platform.claude.com/docs/en/models/sonnet-5-5/overview (checked 2026-10-23).
Released Sep 22 2026. Anthropic's docs recommend it as the starting model for most workloads, built for long-running agentic coding and knowledge work, and point to Fable 5.1 when Opus 5.5 at higher effort still falls short. Adaptive thinking is always on (default effort medium). Priced below Opus 5 ($5/$25). API identifier claude-opus-5-5, reliable knowledge cutoff June 2026, retirement not sooner than Sep 22 2027. Source: https://platform.claude.com/docs/en/models/opus-5-5/overview (checked 2026-09-26).
Released Sep 22 2026 alongside GPT-6 Luna, one tier below GPT-6 Astra in the GPT-6 family. OpenAI priced it at half of GPT-5.6 Sol's promotional $4/$20. API identifier gpt-6-sol, knowledge cutoff Apr 20 2026. Release date per OpenAI's developer community announcement (https://community.openai.com/t/announcing-gpt-6-sol-and-gpt-6-luna-in-the-api-codex-and-chatgpt/1399925).
Released Sep 22 2026 alongside GPT-6 Sol as the high-volume, lowest-cost GPT-6 tier. Replaces GPT-5.6 Luna ($0.20/$1.20) at half the input price. API identifier gpt-6-luna, knowledge cutoff May 18 2026. Release date per OpenAI's developer community announcement (https://community.openai.com/t/announcing-gpt-6-sol-and-gpt-6-luna-in-the-api-codex-and-chatgpt/1399925).
Ships with web search, code interpreter, a hosted shell, apply-patch, computer use, and MCP support built in. Staged rollout starting with a limited partner preview Sep 3 2026, then ChatGPT Plus/Pro/Business/Enterprise plus the API and AWS over the following days. First OpenAI model to hit the 'Critical' cybersecurity threshold under the company's own Preparedness Framework; its most advanced cyber capabilities are gated behind the enterprise-only 'Daybreak' program rather than shipped broadly. Fortune reported OpenAI quietly revised some evaluation metrics upward after a delayed publication of the launch announcement, a credibility flag independent of the ARC-AGI-3 discrepancy above. Strongest, most independently-corroborated advantage across hands-on reviews is cost/token efficiency on agentic coding and robotics tasks rather than raw intelligence; the cross-vendor Artificial Analysis index and Humanity's Last Exam both still favor Claude Fable 5.1 and Opus 5. API identifier gpt-6-astra, knowledge cutoff Apr 30 2026.
Built on the Gemini 3.7 Flash base rather than a new foundation model, per Google's own model card. Tuned specifically for long-horizon coding and autonomous agents. Shipped one day before Claude Fable 5.1 and GPT-6 Astra, in a 48-hour window where three frontier labs released within days of each other; positioned as the cost-efficient option of the three rather than a raw-capability leader. Knowledge cutoff March 2026 (some domains limited to January 2025). Available same-day in GitHub Copilot.
Released Sep 1 2026 alongside Claude Mythos 5.1 (same underlying model, restricted-access safeguards for vetted cybersecurity/life-sciences organizations via Anthropic's Cyber and Life Sciences Verification Programs). Anthropic recommends it over Opus 5 for demanding reasoning and long-horizon agentic work, or when Opus 5 at higher effort still falls short. Anthropic's own comparison shows it beating Fable 5 and Opus 5 on every benchmark listed above except Terminal-Bench 4.0, where sibling model Mythos 5.1 (not Fable 5.1) posts the top score. API identifier claude-fable-5-1, reliable knowledge cutoff June 2026, retirement not sooner than Sep 1 2027.
Same underlying model as Claude Fable 5.1 (identical 1M context, 128K max output, adaptive-always-on thinking, default effort high, Jun 2026 knowledge cutoff, and pricing) but offered separately, by invitation only, as part of Anthropic's Project Glasswing — request access via an Anthropic, AWS, or Google Cloud account team, not available as a general API model. Beats sibling Fable 5.1 on Terminal-Bench 4.0 specifically; the two are otherwise presented as equivalent. Model ID claude-mythos-5-1 (anthropic.claude-mythos-5-1 on Bedrock), retirement not sooner than Sep 1 2027. Source: https://platform.claude.com/docs/en/models/mythos-5-1/overview
Ranked #1 of 170 models on the Artificial Analysis Intelligence Index at 61, ahead of Fable 5 at 60 and GPT-5.6 Sol at 59. Anthropic reports it surpasses all other models on Frontier-Bench, more than doubling Opus 4.8's performance, comes within 0.5 percent of Fable 5's peak CursorBench 3.2 score at half the cost, scores three times the next-best model on ARC-AGI 3, surpasses Fable 5's best OSWorld 2.0 result at about one third of the cost, and posts roughly 1.5 times the next-best pass rate on Zapier AutomationBench. It trails Mythos 5 on cybersecurity.
Select up to 3 models from the directory.
Choose models using their Compare checkboxes to see specifications and costs together.
An LLM leaderboard ranks AI language models side by side using standardised benchmark scores, pricing, speed, and context window size so developers can identify the best model for their workload without reading every provider's documentation.
As of July 2026, Claude Fable 5 leads on SWE-bench Verified (95.0%) and coding tasks, while GPT-5-5 and Gemini 2.5 Pro score competitively on reasoning benchmarks. Use the leaderboard table above and filter by benchmark to see the current rankings.
Select up to three models with the Compare checkboxes, then open the shortlist table. Adjust monthly input and output volumes to calculate the cost of the same workload across all selected models.
Several open-weight models such as Llama 4 Scout and Qwen3 can be accessed for free via third-party providers. Among proprietary APIs, DeepSeek V4 Pro offers strong reasoning at a fraction of the cost of GPT-5 or Claude Fable 5. See the full LLM pricing comparison at /models/pricing.
The leaderboard includes SWE-bench Verified (coding), SWE-bench Pro (harder coding), GPQA Diamond (graduate-level science), AIME (mathematics), OSWorld (computer use), and speed in tokens per second. All scores are sourced from official provider announcements or third-party evaluations.