Explicitly unsupported on this original model; multimodal inputs do not imply media outputs.
Reasoning controls
Use thinking_level; combining it with thinking_budget returns an error. Interactions preserves reasoning history automatically; GenerateContent requires complete history with intact thought signatures.Source for Reasoning controls (opens in a new tab)
These are original documented behaviors. They do not establish API prompt-cache settings.
Access & pricing
Prices belong to an endpoint and billing tier. A provider listing does not guarantee access for every account or region.
Historical identifier; current API calls automatically route to gemini-3.6-flash.
Access
Original release is deprecated and is no longer selectable in Google AI Studio. The identifier remains accepted, but routes to gemini-3.6-flash. Original-checkpoint access through other channels is not established.
Regions
Not collected for the original release
Endpoint pricing has not been collected. Check the provider for applicable rates.
Data & enterprise terms
Retention, training use, regional processing and enterprise terms have not been summarized for this endpoint.
Perplexity Agent API (legacy model selector) · default
Current host-listed selector. Underlying served checkpoint after Google migration is not verified.
Access
Perplexity API key; POST /v1/agent, default service tier. The current host catalog retains this selector and price; original-checkpoint availability is not established.
Current Perplexity host-listed default-tier rates for this legacy selector. Original weights and host-specific token limits after migration are not verified. Tool invocations may add charges. Not Google direct prices. Cached input price: $0.15 / 1M tokens (base tier).
Current Perplexity host-listed default-tier rates for this legacy selector. Original weights and host-specific token limits after migration are not verified. Tool invocations may add charges. Not Google direct prices.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
What the measurements say
Terminal-Bench 2.1
Google DeepMind · provider-reported
76.2percent
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Terminus-2 default harness; self-computed. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. HTML card value conflicts with methodology PDF (53.9). Single attempt; methodology describes five runs using an internal Antigravity harness. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. PDF page 3 conflicts with HTML card (55.1). Five runs, internal Antigravity harness; do not merge the two reports. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Reported by Google from the ScaleAI leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Google attributes results to the benchmark authors at HKUST. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Self-computed; five-run average, single attempt/run, 1080p, at most 100 steps, pyautogui and UI-specific functions. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Google cites the vals.ai public leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Google cites Artificial Analysis. Elo is not a percent. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. No tools; self-computed. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. No tools; mean of Standard (10 options) and Vision settings. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Andon Labs leaderboard score normalized to a 0–100 scale; not unqualified task accuracy. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Cumulative/average 128K result; self-computed. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Pointwise full-context result; not the same aggregation as the 128K score. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. Full text and multimodal set; self-computed. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Original May 2026 evaluation, as reported by Google; not independently reproduced here. ARC Prize verified, semi-private set; Google cites ARC Prize leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-3-5-flash . Default API sampling and pass@1 unless the specific methodology says otherwise.
Latency and throughput depend on the host, request size, reasoning settings and load. There is no measured speed or global usage figure in this record.
The record behind the model
Why it is included
Google documents a distinct original Gemini release with an identified base, launch history, model card and dated evaluations. Distribution is through managed products and APIs; no public weights are offered in these reviewed cards. This is a distribution inference, not an explicit weight license.
