Artificial Analysis: current speed, latency, and version scope

One-sentence takeaway

When Artificial Analysis was reopened on 2026-09-20, its current v4.3.2 snapshot listed Intelligence Index 41/40/33 for high/medium/low. The current page exposed 305 tok/s only for high; current medium and low speed, all TTFT, and low per-task cost remain unknown; high/medium cost $1.24/$0.93 per Index task. The v4.2 figures in the September 2 release material are historical and must not be mixed with the current v4.3.2 snapshot as a regression or speed trend.

Use cases

• Suitable tasks: Coding, Agent, document, and enterprise workflows that require long context, multimodal input, tool calling, and relatively fast answer generation.

• Unsuitable tasks: Using the high tier as the default for low-latency chat, or treating one-off tok/s as total task duration; high has materially higher TTFT and reasoning-token overhead.

• Applicable model version: gemini-3.8-flash; this article records Artificial Analysis's high, medium, and low reasoning pages without treating the three tiers as three API models.

• Applicable client, Agent, or API: The Google API used by Artificial Analysis's tests; Google's official page lists Gemini API, Google AI Studio, Gemini App, Google Antigravity, and other available entry points.

• Recommended reasoning tier and parameters: Start with low for latency-sensitive tasks; test medium for a quality/latency tradeoff; use high for complex, long-horizon Agents. Google's official documentation lists low, medium, and high, and does not support minimal.

Test environment, versions, and measurement timing

• v4.2 release snapshot: The Gemini 3.8 Flash article and X post published by Artificial Analysis on 2026-09-02 used Intelligence Index v4.2; high/medium/low were 59/57/52, respectively. The article also reported about 300 tok/s and 2.5 minutes per task for high, and about 0.8 minutes per task for low.

• v4.3.2 current snapshot: When the Artificial Analysis model page was reopened on 2026-09-20, it showed the current v4.3.2 scope. Only visible Index and high speed values are retained; undisclosed fields remain unknown.

• Performance measurement timing: Artificial Analysis's performance methodology specifies that standard 1k, 10k, and vision workloads are tested eight times per day, with web values shown as the median (P50) over the past 72 hours; the table below is therefore a page snapshot from 2026-09-20, not a fixed permanent value.

• Service and parameters: The primary performance test server is located in Google Cloud us-central1-a; the default temperature for reasoning models is 0.6, and performance benchmarks count tokens with o200k_base. Intelligence Index costs instead use the token counts reported by the provider API.

• Independent/self-reported boundary: Intelligence Index, output speed, TTFT, cost, and task-token figures are independently measured or calculated by Artificial Analysis; Google's model status, modalities, context, and pricing are vendor-reported specifications. Artificial Analysis's statement that the tests are "based on the Google API" does not mean Google endorses its evaluation scores.

Results

Current Artificial Analysis v4.3 model pages

The current Index and high speed above come from the pages reopened on 2026-09-20. Current medium/low speed, all TTFT, and low task cost were not exposed and remain unknown. High/medium cost $1.24/$0.93 per Index task; displayed API input/output prices are $0.75/$3.75 per million tokens. Do not fill the current snapshot with older v4.3 values from 2026-09-08.

Artificial Analysis v4.2 release figures

The v4.2 costs use the release article's definition of “per Intelligence Index task.” The article says high costs about 40% more than Gemini 3.7 Flash's $0.40 because average output tokens increased by about 30% and the number of Agent evaluation turns increased. The v4.2 $0.58 and v4.3 page's $1.24 for high cannot be used to infer a pricing trend because the evaluation set changed from v4.2 to v4.3.

Cross-check against official specifications

Conclusions

1. First distinguish evaluation versions: The v4.2 release figures were high 59, medium 57, and low 52; the 2026-09-20 v4.3.2 page shows 41, 40, and 33. The versions, evaluation sets, and page timestamps differ, so the groups cannot be combined into one ranking or trend.

2. Do not turn unknown values into latency claims: The current page exposes only high speed at 305 tok/s; current medium/low speed, all TTFT, and low task cost remain unknown; high/medium task costs are $1.24/$0.93.

3. Current visible high metrics: v4.3.2 Index is 41 and speed is 305 tok/s; this is not the target client's end-to-end response time.

4. Price cannot be judged only per million tokens: Current input/output prices are both $0.75/$3.75, with a 90% cache-hit discount, but per-task cost is also determined by input, cached, reasoning, and answer tokens plus evaluation turns.

5. The 1M context window is a capability boundary, not a quality guarantee: Both Google and Artificial Analysis list 1M tokens; real long-document workflows still need dedicated testing by input position, retrieval recall, citation accuracy, and end-to-end latency.

Limitations

• The current Artificial Analysis page reopened on 2026-09-20 shows v4.3.2, while the release article uses v4.2; the evaluation sets and timestamps differ, so a “score drop” or cross-version ranking change cannot be calculated directly.

• Performance figures are P50 values over the past 72 hours; TTFT is affected by server location, network conditions, and provider queuing, so results from us-central1-a may not represent mainland China or other regions.

• Artificial Analysis's output speed for reasoning models may be calculated from the latter 80% of the answer chunk, while the performance benchmark and Intelligence Index use different token-counting conventions; tokenizer efficiency limits price comparisons.

• Current medium/low speed, all TTFT, and low per-task cost were not exposed in the 2026-09-20 snapshot; they cannot be inferred from v4.2 or high.

• Evaluations such as AA-Briefcase and GDPval-AA v2 include private test sets or internal scoring processes; the platform's public methodology improves reproducibility, but does not constitute a third-party audit.

• Results reported on the Google DeepMind page for DeepSWE, Vals Finance Agent, Harvey Legal Agent, HLE-Verified, and others are materials published by Google, and this article does not conflate them with independent Artificial Analysis measurements.

Reproduction steps

1. Fix the API model ID at gemini-3.8-flash, run low, medium, and high separately, and record the first reasoning token, first answer token, full response time, input/output/reasoning tokens, and failed retries.

2. For the same set of coding, tool-calling, PDF/long-document, and multimodal tasks, calculate TTFT P50/P95, output tok/s, task completion rate, cost per completed task, and human takeover rate.

3. Repeat the performance measurements across at least two regions or providers, reporting network location, account type, request concurrency, and API version; do not generalize Artificial Analysis's Google API results to other providers.

4. Save the evaluation-set version (v4.2 or v4.3), page collection time, and pricing snapshot; rerun the same task set whenever the model or Artificial Analysis methodology changes.

5. Test 100k-, 500k-, and near-1M-token contexts separately for retrieval, citation accuracy, answer tokens, TTFT, and end-to-end latency, and avoid describing “supports 1M” as “quality remains constant within 1M.”

Original evidence and data

• The current Artificial Analysis pages reopened on 2026-09-20 record v4.3.2 Intelligence Index 41/40/33 for high/medium/low; only high output speed is visible at 305 tok/s, while current medium/low speed, all TTFT, and low task cost remain unknown; high/medium task costs are $1.24/$0.93.

• The Artificial Analysis release article dated 2026-09-02 and its corresponding X post record v4.2 high/medium/low at 59/57/52, costs of $0.58/$0.41/$0.24, approximately 300 tok/s and 2.5 minutes per task for high, and approximately 0.8 minutes per task for low.

• Artificial Analysis's performance methodology states that standard metrics use P50 over the past 72 hours; the 2026-09-20 snapshot should therefore be read as a dated observation, not a permanent value.

• The Google AI Developers model page confirms gemini-3.8-flash, 1,048,576 input tokens, 65,536 output tokens, input modalities, and low/medium/high reasoning tiers; the Google DeepMind model page confirms GA status, 1M input, 64k output, and other product specifications.

Source excerpts or observations (short quotation for compliance only)

The current page exposes high at 305 tok/s but does not expose a current TTFT usable here; output speed still cannot replace user-perceived first-answer delay. Google positions 3.8 Flash as a Flash model for long-horizon software engineering, autonomous Agents, and complex enterprise workflows. Neither source can replace a same-harness rerun for the target business.