Best AI Model for Tool Calling 2026 (September)
By Loïc Jané · Founder, Fleece AI
Best AI Model for Tool Calling in 2026 (September Update)
At a Glance — Updated September 24, 2026.
- Cross-server MCP tool use (MCP-Atlas): Muse Spark 1.1 leads at 88.1%, ahead of Claude Opus 5 (85.8%), Qwen3.8 (84.5%) and GLM-5.3 (84.2%). The top ten sit within about six points, and each score carries a ±2-point margin.
- Large tool catalogs (Toolathlon-Verified, 600+ tools): open-weights models lead. GLM-5.3-Flash scores 78.4% and Kimi K3 76.5%, just ahead of Claude Opus 4.8 at 76.2%.
- Multi-turn tool calls with policy rules (τ³-bench banking): Qwen 3.8 Max leads at 55.2%, followed by Claude Opus 5 (48.7%) and GPT-5.6 Sol (46.9%).
- Long professional tasks (APEX-Agents): Claude Opus 5.5 leads at 73.5% at max effort.
- Not ranked yet: the September releases (Claude Opus 5.5, GPT-6 Astra/Sol/Luna, Gemini 3.8 Flash) are not yet on MCP-Atlas, τ³-bench or Toolathlon-Verified. There is no independent tool-calling score for them yet.
Tool calling (also called function calling) is how an AI model selects the right API, passes valid parameters, reads the result and decides the next step. Every AI agent depends on it. When an agent updates your CRM, files a ticket and posts to Slack in one run, tool calling decides whether that works.
This guide compares current models on the tool-use benchmarks that are still maintained in September 2026. Every score comes from a named leaderboard and a stated date. For what each benchmark measures, see AI Agent Benchmarks 2026 Explained.
Why the February 2026 Numbers No Longer Apply
The first version of this guide (February 2026) ranked GPT-5.2, Claude Opus 4.6 and Gemini 3.1 Pro on τ²-bench, BFCL v4 and MCP-Atlas. Since then, both the models and the benchmarks have changed.
- τ²-bench telecom is saturated. Several models score between 97.8% and 98.2%, so it no longer separates the frontier. Since March 2026, Sierra has moved to τ³-bench, and its new model runs only report the harder banking domain. τ³ also fixed 50+ airline and retail tasks, which raised airline scores by 14 to 20 points. Old and new airline and retail scores cannot be compared.
- BFCL v4 is frozen. The Berkeley Function Calling Leaderboard was last updated on April 12, 2026. Its newest entries are Claude Opus 4.5, GPT-5.2 and Gemini 3 Pro. It cannot rank current models.
- MCP-Atlas changed its method. In April 2026, Scale switched to a new judge and a 100-tool-call budget, and re-scored every model. The February figures (for example, 69.2% for Gemini 3.1 Pro) came from the old method.
- Toolathlon became Toolathlon-Verified. The maintainers say it starts a new score series that cannot be compared with the old board.
The lesson: a tool-calling score is only meaningful with its benchmark version and date attached.
MCP-Atlas: Tool Use Across Real MCP Servers
MCP-Atlas (Scale AI) gives models 1,000 tasks across 36 real MCP servers and 220 tools. A task passes when the answer covers at least 75% of the expected claims.
| Rank | Model | Pass rate |
|---|---|---|
| 1 | Muse Spark 1.1 (Meta) | 88.1% |
| 2 | Claude Opus 5 (xhigh) | 85.8% |
| 3 | Qwen3.8-2.4T-A95B (open weights) | 84.5% |
| 4 | GLM-5.3 (open weights) | 84.2% |
| 5 | Gemini 3.5 Flash (high) | 83.6% |
| 6 | Claude Fable 5 | 83.3% |
| 7 | Kimi K3 (open weights, max) | 82.3% |
| 8 | Claude Opus 4.8 (max) | 82.2% |
| 8 | Muse Spark (Meta) | 82.2% |
| 10 | GPT-5.6 Sol | 81.8% |
| — | Gemini 3.1 Pro Preview (high) | 78.2% |
| — | GPT-5.2 (xhigh) | 67.6% |
Source: Scale AI MCP-Atlas leaderboard, retrieved September 24, 2026.
What it means: most scores carry a margin of about ±2 points, so the top ten are close to a statistical tie. Open-weights models (Qwen3.8, GLM-5.3, Kimi K3) are now level with the proprietary frontier at coordinating MCP tools. That was not true in February.
Toolathlon-Verified: Very Large Tool Catalogs
Toolathlon-Verified runs 108 tasks against 32 MCP servers exposing 604 tools, plus local toolkits. It is the closest public test to an agent connected to hundreds of apps at once.
| Model | Type | Pass@1 |
|---|---|---|
| GLM-5.3-Flash (max) | Open weights | 78.4% |
| Kimi K3 (max) | Open weights | 76.5% |
| Claude Opus 4.8 (max) | Proprietary | 76.2% |
| Muse Spark 1.2 (xhigh) | Proprietary | 75.9% |
| DeepSeek V4 Pro 0813 (max) | Open weights | 74.4% |
| GPT-5.5 (xhigh) | Proprietary | 73.5% |
| Claude Sonnet 5 (max) | Proprietary | 71.6% |
| Gemini 3.5 Flash (high) | Proprietary | 67.3% |
| Gemini 3.1 Pro (high) | Proprietary | 61.1% |
Source: Toolathlon-Verified official leaderboard, retrieved September 24, 2026. Claude Opus 5 and 5.5, Fable 5.1, GPT-5.6, GPT-6 and Gemini 3.8 Flash were not listed.
What it means: when an agent must pick the right tool out of hundreds, the best open-weights models match or beat every proprietary model on this leaderboard. Accuracy still drops as the catalog grows, so loading only the tools a task needs remains good practice.
τ³-bench Banking: Multi-Turn Tool Calls Under Policy
τ³-bench (Sierra) simulates a customer conversation in which the agent must call tools, search a 700-document knowledge base and follow business policy. Sierra now runs new models on the banking domain only, scored as pass^1.
| Model | Pass^1 |
|---|---|
| Qwen 3.8 Max (xhigh) | 55.2% |
| Claude Opus 5 (max) | 48.7% |
| Grok 4.5 (high) | 47.9% |
| GPT-5.6 Sol (xhigh) | 46.9% |
| GPT-5.5 (xhigh) | 44.6% |
| Muse Spark 1.1 (xhigh) | 40.5% |
| Claude Opus 4.7 (max) | 40.2% |
| Kimi K3 (max) | 37.1% |
| GPT-5.2 (high) | 32.2% |
| Gemini 3.1 Pro Preview (high) | 26.0% |
Source: Sierra τ-bench public submissions, retrieved September 24, 2026.
What it means: scores near 50% show how much harder banking is than telecom was. Getting the tool call right is not enough; the agent also has to find the right policy and apply it. For support and operations agents, this is the most realistic tool-calling test available today.
APEX-Agents: Tool Use Inside Long Professional Tasks
APEX-Agents (Mercor) measures end-to-end work by an investment-banking analyst, a management consultant and a corporate lawyer, across files, spreadsheets, chat and calendars.
| Model (effort) | Pass@1 |
|---|---|
| Claude Opus 5.5 (max) | 73.5% |
| Claude Fable 5.1 (max) | 68.6% |
| Gemini 3.7 Flash (high) | 67.8% |
| Claude Opus 5 (max) | 65.8% |
| Grok 4.6 (xhigh) | 65.3% |
Source: Mercor APEX-Agents leaderboard (v1.1), retrieved September 24, 2026. Only the top five are published.
What it means: effort settings matter a lot. Opus 5.5 scores 73.5% at max effort but 52.5% at medium, which is its API default. If you compare models, compare them at the effort level you will actually run.
What About the Newest Models?
Three frontier releases from September 2026 have no independent tool-calling score yet.
- Claude Opus 5.5 (September 22) leads APEX-Agents. Anthropic reports 66.4% on Terminal-Bench 4.0 (vendor-reported, not yet on the public leaderboard).
- GPT-6 Astra (September 3) tops the public Terminal-Bench 4.0 leaderboard at 58.2%. GPT-6 Sol and Luna (September 22) are not listed yet.
- Gemini 3.8 Flash (September 2) scores 19.1% on Terminal-Bench 4.0 (leaderboard). Google has not published MCP-Atlas or τ-bench results for it.
Until these models appear on MCP-Atlas, τ³-bench or Toolathlon-Verified, any claim that one of them is "the best at tool calling" is a guess.
Which Model Should You Use?
The benchmarks point to different winners for different jobs.
- Agents connected to many apps (MCP): Muse Spark 1.1, Claude Opus 5, Qwen3.8, GLM-5.3 and Kimi K3 are within a few points of each other on MCP-Atlas. Choose on price and latency.
- Hundreds of tools in one session: GLM-5.3-Flash and Kimi K3 lead Toolathlon-Verified.
- Support and operations agents that must follow policy: Qwen 3.8 Max leads τ³-bench banking. Claude Opus 5 and GPT-5.6 Sol follow.
- Long, multi-step professional work: Claude Opus 5.5 at high or max effort leads APEX-Agents.
- High-volume, simple tool calls: the cheapest current models (GPT-6 Luna, Gemini 3.8 Flash, Claude Haiku 4.5) have no independent tool-calling score yet. Test them on your own workflows before relying on them.
API Pricing (per 1M Tokens)
Tool-heavy workflows resend the conversation and tool results at every step, so input price matters more than usual.
| Model | Input | Output |
|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 |
| Claude Fable 5.1 | $10.00 | $50.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| GPT-5.6 Sol | $4.00 | $20.00 |
| GPT-6 Sol | $2.00 | $10.00 |
| Claude Sonnet 5 | $2.00 | $10.00 |
| Gemini 3.1 Pro Preview (≤200k context) | $2.00 | $12.00 |
| Grok 4.7 (<200k context) | $2.00 | $6.00 |
| Claude Haiku 4.5 | $1.00 | $5.00 |
| Gemini 3.8 Flash (introductory, until Dec 31, 2026) | $0.75 | $3.75 |
| GPT-6 Luna | $0.10 | $0.50 |
Sources: official pricing pages of OpenAI, Anthropic, Google and xAI, retrieved September 24, 2026. Gemini 3.8 Flash rises to $1.50 / $7.50 on January 1, 2027.
Tool Calling on Fleece AI
Fleece AI lets you choose the model for each agent. Several of the leaders above are available:
- GLM 5.3 Flash, the Fleece default: 82.3% on MCP-Atlas and 76.5% on Toolathlon-Verified.
- GLM 5.3 Flash: #1 on Toolathlon-Verified (78.4%).
- Qwen 3.8 Max: #1 on τ³-bench banking (55.2%).
- Claude Opus 5.5: #1 on APEX-Agents (73.5% at max effort).
- GPT-5.6 Sol: 81.8% on MCP-Atlas and 46.9% on τ³-bench banking.
GPT-6, Gemini and Muse Spark are not offered on Fleece AI today.
Compare them on your own workflows — start your Fleece AI trial and switch the model per agent to see which one calls your apps most reliably.
Frequently Asked Questions
What is the best AI model for tool calling in 2026?
It depends on the benchmark. As of September 24, 2026, Muse Spark 1.1 leads MCP-Atlas (88.1%), GLM-5.3-Flash leads Toolathlon-Verified (78.4%), Qwen 3.8 Max leads τ³-bench banking (55.2%), and Claude Opus 5.5 leads APEX-Agents (73.5%). The newest releases (Opus 5.5, GPT-6, Gemini 3.8 Flash) are not yet on the three dedicated tool-use leaderboards.
Is τ²-bench still a good tool-calling benchmark?
Not for ranking frontier models. The telecom domain is saturated, with several models between 97.8% and 98.2%. Sierra now reports new models on the harder τ³-bench banking domain, where the best score is 55.2%.
Is BFCL still updated?
No. The Berkeley Function Calling Leaderboard (BFCL v4) was last updated on April 12, 2026. Its newest models are Claude Opus 4.5, GPT-5.2 and Gemini 3 Pro, so it cannot rank models released since.
Are open-weights models good at tool calling?
Yes. GLM-5.3-Flash and Kimi K3 lead Toolathlon-Verified, and Qwen3.8, GLM-5.3 and Kimi K3 are within six points of the leader on MCP-Atlas, where most scores carry a ±2-point margin.
Does tool calling accuracy drop with more tools?
Yes. Accuracy falls as the number of available tools grows, which is what Toolathlon-Verified's 600+ tool setup tests. Loading only the tools relevant to the task, rather than every tool at once, keeps agents more reliable.
Related Articles
- AI Agent Benchmarks 2026 Explained — what each benchmark measures and who leads
- Best AI Models for Workflow Automation 2026 — full model comparison
- Model Context Protocol (MCP) Explained — the standard behind MCP-Atlas
Start automating with AI agents — deploy your first AI agent with Fleece AI and pick the model that fits each job.