Multimodal reasoning model for coding, scientific work and agent workflows. Input, thinking and output share one token budget; extended output uses continuation mode.
Upper shared ceiling; actual available thinking/output is 1,048,576 minus input.
Long output mode
Interactions: background=true, store=true and max_output_tokens > 65,536. Server continues in hops up to 64K, for up to 24 hours. Without an explicit larger budget, background runs still stop at 64K. GenerateContent supports client-side continuation with 64K streamed or 32K non-streamed per-request output and a 20-minute run limit.Source for Long output mode (opens in a new tab)
Each hop replays preceding content as input and is billed. Interactions client-side continuation is described as coming soon, not currently available.
Multimodal pretraining uses web material, text/code, images, audio and video. Sources include public datasets, crawled and licensed material, synthetic content, workforce/business data and user data subject to applicable controls. Post-training uses instruction, reinforcement-learning and preference data.Source for Published training data (opens in a new tab)
Paid Gemini API / AI Studio account. standard consumption option; account, quota and region restrictions apply. Standard Interactions synchronous output is limited to 65,536 tokens. Long decode requires background execution with store=true and a larger output budget.
Regions
Model-specific country/account eligibility not collected
standard introductory rates through November 30, 2026. From December 1: input $4, output $20, cached input $0.2 per million tokens. Output includes thinking. No input-length price tier is published here. Long decode additionally bills replayed input per hop; grounding can add charges. Cached input price: $0.10 / 1M tokens (base tier).
Published billing units · original currency · observed 14 Oct 2026
Category
Usage
Rate
Unit
Scheduled December 2026 pricing
Input1
$4.00
1M tokens
Scheduled December 2026 pricing
Output including thinking1
$20.00
1M tokens
Scheduled December 2026 pricing
Cached input1
$0.20
1M tokens
Grounding add-on
Google Search2
$14.00
1K search queries
Grounding add-on
Google Maps2
$14.00
1K search queries
Rate conditions
Announced rates effective December 1, 2026. Not the current October 14 quote. Output includes thinking; tool charges are separate.
After the published 5,000 monthly free allowance shared across Gemini 3 and newer models for this grounding product. A single user request can trigger multiple billable search queries; not $14 per 1,000 questions.
Paid-tier pricing table states submitted data is not used to improve products. This is not a zero-retention claim; long-decode background interactions require server storage.
Paid Gemini API / AI Studio account. batch consumption option; account, quota and region restrictions apply. Standard Interactions synchronous output is limited to 65,536 tokens. Long decode requires background execution with store=true and a larger output budget.
Regions
Model-specific country/account eligibility not collected
Streaming behavior for this specific billing tier is not established by the generic long-decode guide.
INPUT / 1M TOKENS$1.00
OUTPUT / 1M TOKENS$5.00
batch introductory rates through November 30, 2026. From December 1: input $2, output $10, cached input $0.1 per million tokens. Output includes thinking. No input-length price tier is published here. Long decode additionally bills replayed input per hop; grounding can add charges. Cached input price: $0.05 / 1M tokens (base tier).
Paid-tier pricing table states submitted data is not used to improve products. This is not a zero-retention claim; long-decode background interactions require server storage.
Paid Gemini API / AI Studio account. flex consumption option; account, quota and region restrictions apply. Standard Interactions synchronous output is limited to 65,536 tokens. Long decode requires background execution with store=true and a larger output budget.
Regions
Model-specific country/account eligibility not collected
Streaming behavior for this specific billing tier is not established by the generic long-decode guide.
INPUT / 1M TOKENS$1.00
OUTPUT / 1M TOKENS$5.00
flex introductory rates through November 30, 2026. From December 1: input $2, output $10, cached input $0.1 per million tokens. Output includes thinking. No input-length price tier is published here. Long decode additionally bills replayed input per hop; grounding can add charges. Cached input price: $0.05 / 1M tokens (base tier).
Paid-tier pricing table states submitted data is not used to improve products. This is not a zero-retention claim; long-decode background interactions require server storage.
Paid Gemini API / AI Studio account. priority consumption option; account, quota and region restrictions apply. Standard Interactions synchronous output is limited to 65,536 tokens. Long decode requires background execution with store=true and a larger output budget.
Regions
Model-specific country/account eligibility not collected
Streaming behavior for this specific billing tier is not established by the generic long-decode guide.
INPUT / 1M TOKENS$3.60
OUTPUT / 1M TOKENS$18.00
priority introductory rates through November 30, 2026. From December 1: input $7.2, output $36, cached input $0.36 per million tokens. Output includes thinking. No input-length price tier is published here. Long decode additionally bills replayed input per hop; grounding can add charges. Cached input price: $0.18 / 1M tokens (base tier).
Published billing units · original currency · observed 14 Oct 2026
Category
Usage
Rate
Unit
Scheduled December 2026 pricing
Input1
$7.20
1M tokens
Scheduled December 2026 pricing
Output including thinking1
$36.00
1M tokens
Scheduled December 2026 pricing
Cached input1
$0.36
1M tokens
Grounding add-on
Google Search2
$14.00
1K search queries
Grounding add-on
Google Maps2
$14.00
1K search queries
Rate conditions
Announced rates effective December 1, 2026. Not the current October 14 quote. Output includes thinking; tool charges are separate.
After the published 5,000 monthly free allowance shared across Gemini 3 and newer models for this grounding product. A single user request can trigger multiple billable search queries; not $14 per 1,000 questions.
Paid-tier pricing table states submitted data is not used to improve products. This is not a zero-retention claim; long-decode background interactions require server storage.
Conditional effective standard rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $4 input, $20 output, $0.2 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.10 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Conditional effective standard rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $4.4 input, $22 output, $0.22 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.11 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Conditional effective standard rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $4.4 input, $22 output, $0.22 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.11 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Conditional effective priority rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $7.2 input, $36 output, $0.36 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.18 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Conditional effective priority rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $7.92 input, $39.6 output, $0.396 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.198 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Conditional effective priority rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $7.92 input, $39.6 output, $0.396 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.198 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Conditional effective batch rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $2 input, $10 output, $0.1 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.05 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Conditional effective batch rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $2.2 input, $11 output, $0.11 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.055 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Conditional effective batch rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $2.2 input, $11 output, $0.11 cached input per million. Output includes thinking. Long decode can bill replayed input. Cached input price: $0.055 / 1M tokens (base tier).
Project-scoped implicit cache can retain data up to 24 hours. Durable cache is enabled by default and does not support CMEK. Regional cached data stays in the selected location; global data may be served from any region. Abuse monitoring is governed separately by account/model terms.
Implicit-cache TTL is not published in the reviewed guide.
How to enable it
No cache-object creation or cached_content parameter. Explicit caching is unsupported. Common prompt prefixes and nearby requests increase hit probability.Source for How to enable it (opens in a new tab)
Usage fields
Interactions usage.total_cached_tokens reports hits; long decode reports cumulative usage across hops. Read the latest totals rather than summing streaming snapshots.Source for Usage fields (opens in a new tab)
Billing rules
Cached reads receive a 95% input discount at the applicable tier rate. Long decode replays prior content each hop, with repeated input discounted by implicit caching.Source for Billing rules (opens in a new tab)
standard introductory rates through November 30, 2026. From December 1: input $4, output $20, cached input $0.2 per million tokens. Output includes thinking. No input-length price tier is published here. Long decode additionally bills replayed input per hop; grounding can add charges.
Announced rates effective December 1, 2026. Not the current October 14 quote. Output includes thinking; tool charges are separate.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Implicit-cache TTL is not published in the reviewed guide.
How to enable it
No cache-object creation or cached_content parameter. Explicit caching is unsupported. Common prompt prefixes and nearby requests increase hit probability.Source for How to enable it (opens in a new tab)
Usage fields
Interactions usage.total_cached_tokens reports hits; long decode reports cumulative usage across hops. Read the latest totals rather than summing streaming snapshots.Source for Usage fields (opens in a new tab)
Billing rules
Cached reads receive a 95% input discount at the applicable tier rate. Long decode replays prior content each hop, with repeated input discounted by implicit caching.Source for Billing rules (opens in a new tab)
batch introductory rates through November 30, 2026. From December 1: input $2, output $10, cached input $0.1 per million tokens. Output includes thinking. No input-length price tier is published here. Long decode additionally bills replayed input per hop; grounding can add charges.
Announced rates effective December 1, 2026. Not the current October 14 quote. Output includes thinking; tool charges are separate.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Implicit-cache TTL is not published in the reviewed guide.
How to enable it
No cache-object creation or cached_content parameter. Explicit caching is unsupported. Common prompt prefixes and nearby requests increase hit probability.Source for How to enable it (opens in a new tab)
Usage fields
Interactions usage.total_cached_tokens reports hits; long decode reports cumulative usage across hops. Read the latest totals rather than summing streaming snapshots.Source for Usage fields (opens in a new tab)
Billing rules
Cached reads receive a 95% input discount at the applicable tier rate. Long decode replays prior content each hop, with repeated input discounted by implicit caching.Source for Billing rules (opens in a new tab)
flex introductory rates through November 30, 2026. From December 1: input $2, output $10, cached input $0.1 per million tokens. Output includes thinking. No input-length price tier is published here. Long decode additionally bills replayed input per hop; grounding can add charges.
Announced rates effective December 1, 2026. Not the current October 14 quote. Output includes thinking; tool charges are separate.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Implicit-cache TTL is not published in the reviewed guide.
How to enable it
No cache-object creation or cached_content parameter. Explicit caching is unsupported. Common prompt prefixes and nearby requests increase hit probability.Source for How to enable it (opens in a new tab)
Usage fields
Interactions usage.total_cached_tokens reports hits; long decode reports cumulative usage across hops. Read the latest totals rather than summing streaming snapshots.Source for Usage fields (opens in a new tab)
Billing rules
Cached reads receive a 95% input discount at the applicable tier rate. Long decode replays prior content each hop, with repeated input discounted by implicit caching.Source for Billing rules (opens in a new tab)
priority introductory rates through November 30, 2026. From December 1: input $7.2, output $36, cached input $0.36 per million tokens. Output includes thinking. No input-length price tier is published here. Long decode additionally bills replayed input per hop; grounding can add charges.
Announced rates effective December 1, 2026. Not the current October 14 quote. Output includes thinking; tool charges are separate.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (standard, global) · standard
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective standard rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $4 input, $20 output, $0.2 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (standard, us) · standard
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective standard rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $4.4 input, $22 output, $0.22 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (standard, eu) · standard
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective standard rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $4.4 input, $22 output, $0.22 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (priority, global) · priority
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective priority rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $7.2 input, $36 output, $0.36 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (priority, us) · priority
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective priority rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $7.92 input, $39.6 output, $0.396 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (priority, eu) · priority
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective priority rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $7.92 input, $39.6 output, $0.396 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (batch, global) · batch
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective batch rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $2 input, $10 output, $0.1 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (batch, us) · batch
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective batch rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $2.2 input, $11 output, $0.11 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
Observed 14 Oct 2026. Write rates and storage fees are different charges; use the billing rules above.
Caching reviewed 14 Oct 2026
Google Enterprise Agent Platform (batch, eu) · batch
The generic cache guide instead gives 4,096 for Gemini 4. This host-specific discrepancy is unresolved; the model card is displayed. Developer API has a separate 4,096-token statement.
Common prompt prefixes may hit automatically. Disable durable caching alone or all implicit caching per project; explicit cache objects are unsupported.Source for How to enable it (opens in a new tab)
Published Argon token rates give a 95% cached-read discount, despite the generic guide describing 90%. Implicit storage has no separate charge.Source for Billing rules (opens in a new tab)
Limits & data boundaries
Durable KV data is stored at rest using Google encryption, not CMEK; project-isolated, never used for training, retained at most 24 hours. Regional endpoints keep cached data in the selected location; global requests/cache may be served from any region.Source for Limits & data boundaries (opens in a new tab)
Conditional effective batch rates after 50% credits back on net spend through November 30, 2026, not an unconditional invoice rate. Published rates from December 1 are $2.2 input, $11 output, $0.11 cached input per million. Output includes thinking. Long decode can bill replayed input.
Published full rates from December 1, 2026. Current introductory values are delivered via 50% credits back on net spend; no quoteRole is assigned.
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
Vals Index
Google DeepMind · provider-reported
68.9percent
October 2026 model-card report; independently reproducing results is outside this record. Google cites Vals AI. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Private set; Google cites Zapier public leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Google cites Vals AI. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Google cites Vals AI. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Self-computed mini-SWE agent harness. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Google cites Proximal leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Google cites Vals AI leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Self-computed; highest scoring thinking level. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Self-computed OpenCode harness, 10-hour budget on one NVIDIA H100; weighted aggregate over four base models and seven benchmarks. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Self-computed; verifier timeout increased sixfold. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Linux terminal, bioinformatics tools, Python/R and internet access. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Google cites Surge leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. BFS, 650-item subset with context lengths up to 128K. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. BFS, distinct 200-problem subset with 256K to 1M contexts. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Default ALE-Claw harness, five-hour window, safety filters enabled; binary pass rate. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Partial score, best of three single-attempt runs. 1080p, 500-step limit, parallel UI tool calls, compaction, screenshot-only observations; August 8 patch applied. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. No tools; Google cites Surge public leaderboard. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Self-computed, no tools, Gemini video at one frame per second. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
October 2026 model-card report; independently reproducing results is outside this record. Google cites official leaderboard. Score is pass@1; leaderboard ties are broken using pass@4. Methodology: https://deepmind.google/models/evals-methodology/gemini-4-argon . Highest thinking setting and pass@1 unless a benchmark-specific exception is stated. The methodology describes benchmark coverage as September 2026 while the result table is labeled October 2026.
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 describes a new original model, not a fine-tune of an earlier Gemini release, distributed through proprietary products and managed APIs. The reviewed official release supplies no downloadable weights; this is a distribution inference rather than an explicit weight-license declaration.
