Google Gemini Decisions API: pricing, rivals and limits | Magica
Google launches Gemini Decisions API to cut the cost of AI workflow choices
Editorial Team
••📖7 min read
Google’s Gemini Decision 1 scores predefined choices across text, images, audio, video and PDFs for $0.10 per million input tokens. Its shared-input billing targets recurring workflow costs, but cheaper decision models, rival caching and low-cost generative alternatives complicate the savings claim.
Gemini Decision 1 scores up to 1,000 independent questions over shared text or media, returning probabilities rather than prose.
Google charges $0.10 per million input tokens and bills shared input once within a request; repeated requests pay for that input again.
Microsoft advertises a lower decision-model input rate, while OpenAI’s competing endpoint supports caching across requests.
Google launched the Gemini Decisions API on November 11, 2026, offering developers a specialized model for choosing among predefined options without generating answer text. Gemini Decision 1 accepts text, images, audio, video and PDFs, with a price of $0.10 per million input tokens and no output-token charge, Google said in its announcement.
The target is a recurring expense inside AI applications: deciding which model should handle a request, which tool to select, whether a policy applies or how relevant a retrieved document is. Moving those judgments to a specialized scorer could reduce the time and money spent before a larger model produces an answer. Google says its own teams already use the service for routing, design and safety checks.
The launch adds a multimodal option to an existing competitive category. Microsoft released a single-pass decision model in October at a lower advertised input-token rate. OpenAI’s decision endpoint is already generally available and offers caching that Google’s new endpoint lacks. The useful comparison is the cost of completing a workflow accurately, including any follow-up calls and review.
Google’s illustrative demonstration shows company-reported outputs for spill detection, cleanup selection and severity scoring. Source: Google.
A narrower job than Google’s generative models
Gemini Decision 1 extends Google’s Gemini model family with a purpose-built scorer. Its specification describes a single prefill pass: the model processes the supplied evidence and questions without decoding a text response token by token. It returns probabilities and confidence values that application code can use directly.
Palantir’s co-founder and CEO says he could back a Democrat in 2028, while excluding a Democratic Socialists of America member. His Democratic history complicates talk of a political conversion; the more immediate question is how his government software company will answer congressional scrutiny.
Developers can ask yes/no questions, choose among two to 255 options, or rate an input against ordered rubric levels. A rubric score is the probability-weighted average of the level indices, so it can fall between categories. Up to 1,000 independent questions can share a request, within a 262,144-token context window covering both the input and all questions.
Google already offered alternatives for programs that need structured answers. Its release history records JSON-schema support for Gemini 1.5 Flash in August 2024 and the July 2026 general-availability release of Gemini 3.5 Flash-Lite, a generative model designed for high-volume automation. The novelty here is a dedicated probability-scoring interface, rather than the first way to make Gemini return machine-readable results.
The Decisions API guide directs developers to Google’s Interactions API for generated explanations, extracted objects, tool calls with arguments or multi-step reasoning. Decision 1 can score which tool fits a task, but does not support function calling: the application must execute the selected action.
Company-reported input rates, charted by Magica. Haiku 5.5’s rate covers prompts up to 100,000 tokens; output charges and caching discounts are excluded. These are not matched tests of workflow cost. Source: Microsoft.
Shared-input billing is the economic lever
Google bills the shared input once per request, regardless of how many questions examine it. Each question adds its own instruction and criteria tokens. The company says there are no separate thinking-token, audio-transcription or optical-character-recognition charges, and offers a no-cost tier for prototypes in Google AI Studio, its developer environment.
That structure favors several independent checks over the same document or media input. Image and video resolution settings let developers trade visual detail against token cost. But the endpoint does not support context caching, Batch API, Flex inference or Priority inference; input sent again in another request is billed in full.
Google says Decision 1 through the Decisions API is nearly five times cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite through the Interactions API. That is a company-reported comparison with one generative alternative. A separate account of Google’s comparison says the Flash-Lite calculation also assumed zero billed output tokens, underscoring that the claimed saving cannot simply be equated with removing an output bill.
Microsoft’s October 9 announcement describes Microsoft-Decision-1 as a specialized scorer available through Microsoft Foundry, its model platform, and OpenRouter. Microsoft says it built the model by post-training the existing Qwen3.5-9B model for single-pass scoring. Like Google’s model, it supports yes/no, multiple-choice and rating options with calibrated probabilities. Its advertised price is $0.042 per million input tokens, with free output tokens, compared with Google’s $0.10 input rate.
OpenAI’s October 22 release notes say its Decisions API became generally available with GPT-6 Luna, the model powering that endpoint. It supports up to 400 questions per request and caching of shared input across requests for discounts. Google’s higher question ceiling therefore comes with a different reuse trade-off: developers must consider how many checks they can group together and how often they resend the same evidence.
Small generative models remain substitutes where a workflow also needs written output. Anthropic’s October 7 launch positioned Claude Haiku 5.5 for high-volume tasks including classification and summaries. For prompts up to 100,000 tokens, its listed price is $0.10 per million input tokens—the same base input rate as Decision 1—and $0.50 per million output tokens. Longer prompts have higher rates.
These are advertised token prices and feature limits, rather than matched tests of cost per correct decision. Token counts, supported inputs, errors and fallback calls can change the total. A dedicated scorer must earn its place against both other decision endpoints and a cheap generative call that may do more of the job.
Speed claims stop short of the maximum request
Google reports median model-compute latency of 50–100 milliseconds and end-to-end latency of 150–250 milliseconds across multimodal inputs with contexts up to 16K tokens. It says latency stayed nearly flat from one to 100 questions per request. Those measurements do not establish equivalent speed at the full context limit or the maximum 1,000 questions.
The company also claims leading performance and the lowest calibration error across its benchmark suite for routing, reasoning, ranking and multimodal understanding. These remain Google’s measurements. Calibration describes how closely confidence estimates track actual correctness; a high-confidence answer can still be wrong.
Google says Gemini Notebook uses the API to route requests among model tiers and thinking levels, reducing latency for everyday queries by up to 40%. Its other examples include Stitch’s rapid design experiment, emoji selection in CC from Google Labs, and screening for adversarial content and behavior around tools in Google Search. These are company-described internal uses, rather than evidence that outside developers will achieve the same gains.
Access is initially through the Gemini Developer API’s v1beta channel with Decision 1; Google lists support through Gemini in Google Cloud as coming soon. The quickstarts require version 3.2 or later of Google’s Gen AI SDK. Conversation history and tool-call blocks cannot be passed directly as input blocks, although multi-turn context can be serialized as text.
Vercel’s AI Gateway, which provides unified model access and tracks usage and costs, has also added Decision 1. Its integration accepts PDF, audio and video through AI SDK 7, while its TypeSafe-compatible and Decisions APIs support text and images. The route a developer chooses therefore affects which inputs are available.
Google’s diagram shows shared input feeding independent questions, with a typed answer and confidence value for each. Source: Google Decisions API guide.
The next test is the cost of acting correctly
Google’s guide says each question is evaluated independently against the shared input: questions cannot see one another’s answers. A dependent decision requires another request, or asking both questions together and discarding an answer the application does not need. Sequential workflows cannot assume that the maximum question count translates into equivalent parallel savings.
A choice question also distributes probability among the supplied options even when none fits. Google recommends confidence thresholds, a separate scope check or an explicit catch-all option. Thresholds should be set using labeled examples from actual traffic, with the cost of false positives weighed against false negatives. Its guidance also calls for adversarial testing and a human or second check for consequential actions such as payouts.
The next decision for developers is whether representative traffic produces enough correct, usable answers to justify inserting another model call. A useful evaluation must measure accuracy, calibration, tail latency and the share of requests needing a retry, a larger model or human review, alongside the full bill. Those results will determine whether shared-input scoring improves the application’s economics.
ByteDance’s video model is reaching creators through multiple platforms with 30-second 4K output. Posted prices and draft workflows show why the choice of distributor still matters.
Rox’s $1.3 billion valuation backs AI agents for large enterprise sales teams. Its growth and cost claims face a crowded market, while integration with existing systems leaves the case for replacing Salesforce unresolved.
Treasury and other agencies reportedly want Singapore cloud provider Megaspeed on a U.S. trade blacklist. Commerce denies a reported cabinet confrontation over the delay, while Nvidia says it has blocked shipments since June 2025.
United is restricting unauthorized AI agents from searching, booking and managing travel through its platforms and intermediaries. The airline cites security and safety risks as it reserves control over future integrations.
A controlled study found that pharmaceutical ads shifted AI recommendations toward promoted drugs without reducing overall clinical accuracy. The results expose a gap in accuracy tests, but do not establish effects on prescriptions or patient outcomes.
Former AWS computing chief Dave Brown will succeed Santosh Janardhan in early 2027 as Meta weighs its next capacity budget after a quarter of negative free cash flow.
Google is rolling out its Gemini app on Apple CarPlay, bringing voice conversations linked to Gmail and Google Calendar. It joins existing in-car AI options, while some personalized features remain restricted by age and region.
Google Labs’ Tabletop lets groups generate and use shared itineraries, polls and other interactive tools. The US adult-only experiment builds on an idea already offered by shareable AI app builders, leaving its practical advantage to be tested.
Anthropic’s IPO has reportedly attracted orders multiple times the available shares. Its planned $3 billion–$3.5 billion Japanese allocation would exceed Japan’s reported domestic IPO proceeds so far this year, but indications of interest are not completed sales.
Google warned that charging for news links would force it to stop distributing Indonesian news, while Meta and TikTok challenged other parts of the copyright bill. Existing Indonesian payment mechanisms and overseas deals complicate the platforms’ warnings about the cost to publishers.
Utrecht is investigating restrictions on camera glasses in municipal buildings, with findings expected in early 2027. The city must establish whether a ban is lawful, enforceable and compatible with assistive uses.
Research commissioned by cloud-storage supplier Wasabi estimates £3.8 billion in annual UK infrastructure overspend attributable to hard-to-forecast provider charges, including £1.2 billion for AI. Businesses report hiring and investment trade-offs, but recovering that money and turning it into jobs remain conditional.
The USDT issuer’s 800-million-parameter BotPsy beats a larger Qwen model on five of six company-reported agent benchmarks. Its local-AI ambitions still face missing handset measurements and explicit limits on sensitive uses.
Google DeepMind’s AlphaGenome 2 improves cell-specific DNA predictions and accepts new single-cell data without fine-tuning. Company benchmarks show gains at high confidence, but limited recall, complementary tools and greater computational demands leave experimental validation essential.
Coinbase is rolling out server-enforced trading limits and portfolio-loss triggers for AI agents connected to customer accounts. Product-category permissions remain forthcoming, and pausing new trades does not establish a ceiling on losses.
JetBrains IDEs 2026.3 bundle Air for running coding agents and reviewing their changes, while the standalone Air app is in maintenance mode. Existing provider subscriptions remain usable, but billing, security controls and execution options vary by integration.
Amazon’s shares and convertible notes could be worth $271.8 billion at the top of Anthropic’s proposed IPO range. Existing accounting gains, a 12% ownership cap and Anthropic’s computing commitments shape what that paper value means.
Nvidia chief Jensen Huang argued in May 2025 that US chip restrictions were strengthening Chinese rivals. Filings show a substantial commercial stake, but lost sales and market share alone cannot establish whether the controls failed their security purpose.
Reported demand has Anthropic’s bankers considering a price above its filed $70–$78 range. The combined financing could exceed $100 billion, but existing cash, uneven operating results and long-term computing commitments complicate the valuation test.
