Gemini 3.0 Pro scores 85.4% on τ2-bench tool use benchmark, up from 54.9% in version 2.5. The model won 19 out of 20 benchmarks against GPT-5.1 and Claude Sonnet 4.5.

The 9.5x performance jump on long-horizon tasks comes from Thought Signatures—encrypted strings that preserve reasoning between function calls. When an agent performs 50+ operations (analyzing invoices, comparing prices, negotiating terms), each signature maintains why previous steps matter. Without them, context is lost after a few calls.

τ2-bench (tool use): 85.4% vs 54.9% in 2.5
Terminal coding: 54.2% vs 32.6%
SWE-Bench: 76.2% vs 59.6%
Vending-Bench: $5,478 vs $574 (9.5x improvement)
MathArena: 23.4% vs 0.5% (46x improvement)

Thought Signatures are mandatory in Gemini 3.0 Pro for all tool calls. Every function call requires one, and they must be preserved exactly as received. Missing a signature results in an HTTP 400 error that breaks the entire workflow.

Platforms that implement tool calling via text responses and custom parsing (to stay LLM-agnostic and work with all models) must also preserve these signatures. These text-based tool calls are still tool calls—they just use custom parsing instead of native function calling. Gemini stores its reasoning for these commands in the signatures, so they must be maintained.

Example with native function calling:
{
"functionCall": {
"name": "analyze_invoice",
"args": {"id": "INV-2024-001"}
},
"thoughtSignature": "<encrypted_execution_context>"
}

Example with text-based tool calling (LLM-agnostic):
TOOL_CALL: analyze_invoice {"id": "INV-2024-001"}
THOUGHT_SIGNATURE: <encrypted_execution_context>
# Still a tool call, just parsed from text

Parallel calls (single signature):
[
{"functionCall": {"name": "get_temp", "city": "Paris"},
"thoughtSignature": "sig_A"},
{"functionCall": {"name": "get_temp", "city": "London"}}
]

Sequential calls (unique signatures):
Step 1: {"functionCall": {"name": "check"}, "thoughtSignature": "sig_1"}
Step 2: {"functionCall": {"name": "act"}, "thoughtSignature": "sig_2"}

Each signature must be preserved in conversation history.

Gemini 3.0 Pro makes Thought Signatures mandatory. Every platform—n8n, Make, Zapier—must integrate this API or agents fail with HTTP 400 errors.

Thought Signatures preserve the chain of thought across operations. A pricing anomaly in invoice #3 connects to patterns in invoice #47. Without signatures, these connections are lost.

OpenKBS implements full Thought Signature persistence. Every signature is stored, validated, and passed correctly through parallel and sequential operations.

Gemini 3.0 Pro is available in OpenKBS today with full Thought Signature support. The platform handles signature persistence automatically in the agent conversation history.

The 85.4% tool use accuracy enables building invoice processing agents that catch pricing anomalies or customer service agents that actually resolve tickets.

Available at openkbs.com