Google Launches Tool Retrieval for Gemini API Agents
| Parameter | Before | Now |
|---|---|---|
| Tool library management | Full tool schemas sent in every prompt | Dynamic on-demand loading of schemas |
What exactly changed in AI models: context, prices, capabilities. Only values stated by sources — never guessed.
| Parameter | Before | Now |
|---|---|---|
| Tool library management | Full tool schemas sent in every prompt | Dynamic on-demand loading of schemas |
| Parameter | Before | Now |
|---|---|---|
| Terminal-Bench 4.1 pass@1 | 5.6% | 12.0% |
| Context window | 256K | 256K |
| Parameter | Before | Now |
|---|---|---|
| Audio input format | Not specified | 16-bit PCM mono (16 or 24 kHz) |
| Max connection duration | Not specified | 30 minutes |
| Language support | Not specified | English only |
| Parameter | Before | Now |
|---|---|---|
| Context window | Not specified | 1,000,000 tokens |
| Input price | Not specified | $3.00 per 1M tokens |
| Output price | Not specified | $15.00 per 1M tokens |
| Parameter | Before | Now |
|---|---|---|
| Context window | Not specified | 1,048,576 tokens |
| Input price | Not specified | $0.30 per 1M tokens |
| Output price | Not specified | $1.20 per 1M tokens |
| Parameter | Before | Now |
|---|---|---|
| Context window | Not specified | 1 million tokens |
| Input price | Not specified | $3 per million tokens |
| Cached input price | Not specified | $0.15 per million tokens |
| Output price | Not specified | $15 per million tokens |
| Parameter | Before | Now |
|---|---|---|
| Parameters | Not specified | 1B |
| Context window | Not specified | 131k |
| Training tokens | Not specified | 237 billion |
| Parameter | Before | Now |
|---|---|---|
| Session duration | 60 seconds (Genie 3) | 3 minutes (Genie 3) |
| Audio capability | Not specified | Real-time native adaptive audio |
| Grounding | Not specified | 360-degree panorama image grounding |
| Parameter | Before | Now |
|---|---|---|
| Parameter count | Not specified | 0.5B |
| Throughput | Not specified | 7,000+ RTFx |
| Parameter | Before | Now |
|---|---|---|
| Total parameters | Not specified | 5.24 billion |
| Memory requirement | Not specified | 10.5 GB GPU memory (bfloat16) |
| Parameter | Before | Now |
|---|---|---|
| Serving mode | Not specified | Aggregated and disaggregated |
| Routing | Not specified | KV-aware |
| Deployment | Not specified | Multi-node via KAI Scheduler |
| Parameter | Before | Now |
|---|---|---|
| Routing strategy | RANDOM, LEAST_OUTSTANDING_REQUESTS, PREFIX_AWARE | RANDOM, LEAST_OUTSTANDING_REQUESTS, PREFIX_AWARE, KV_AWARE |
| Parameter | Before | Now |
|---|---|---|
| MAI-Cyber-2 | Not specified | Preview |
| MAI Cyber 1.1 Flash | Not specified | Preview |
| Parameter | Before | Now |
|---|---|---|
| Task Execution | Passive assistant | Proactive multi-step agent |
| Parameter | Before | Now |
|---|---|---|
| Base Model | Not specified | Nemotron 3.5 Super VL |
| Parameter | Before | Now |
|---|---|---|
| Task handling | Specialist model cascades | Unified audio-to-JSON inference |
| Parameter | Before | Now |
|---|---|---|
| Supported languages | 18 | 32 |
| Diarization Error Rate (DER) | Not specified | 29% lower |
| Price | $0.21/hr | $0.21/hr |
| Parameter | Before | Now |
|---|---|---|
| Accuracy focus | Standard STT optimization | Optimized for quiet user turns and background noise rejection |
| Entity handling | Not specified | Improved performance on numbers and turn-end detection |
| Parameter | Before | Now |
|---|---|---|
| Model architecture | Not specified | Specialized variants for recording, realtime, and voice agents |
| Parameter | Before | Now |
|---|---|---|
| Training performance per GPU | GB300 NVL72 | Up to 2x improvement with Vera Rubin NVL72 |
| Agentic AI throughput per watt | GB300 NVL72 | >10x improvement with Vera Rubin NVL72 |
| Parameter | Before | Now |
|---|---|---|
| Context window | Not specified | 98304 tokens |
| Speculative decoding | Not specified | Multi-token prediction (4 tokens) |
| KV-cache | Not specified | FP8 storage |
| Parameter | Before | Now |
|---|---|---|
| Web search | Not supported | Supported via deepseek_web_search tool |
| File input | Not supported | PDF, Office, image parsing |
| URL reading | Not supported | Supported via deepseek_read_url tool |
| Safety check | Not supported | Optional parameter for risk control |
| Parameter | Before | Now |
|---|---|---|
| Capability | Not specified | Model reasoning about assembly and constraint generation |
| Parameter | Before | Now |
|---|---|---|
| Architecture | Standard Granite 4.1 Instruct | LLMON-native structure with special-token delimiters |
| Context Handling | Flat text/JSON | Addressable blocks (instruction, data, requirement, exec) |