Gemini is Google’s general AI assistant and model family across consumer, developer, and enterprise products. Gemini 3.7 Flash strengthens the fast-model tier for software engineering, web development, knowledge work, and agent workflows.
OUR VERDICT
A consequential Flash upgrade, especially for agent builders
The combination of stronger tool use and temporary introductory pricing makes 3.7 Flash worth a controlled evaluation. Buyers should budget for the scheduled 2027 price and validate Google’s benchmark claims on their own tasks.
What Gemini 3.7 Flash changes
Google positions 3.7 Flash as its most capable workhorse model for coding and agents. The company reports improvements in debugging, issue resolution, first-pass code, web application generation, complex documents, business automation, instruction following, and recovery from roadblocks.
Availability
Developers can access the model through the Gemini API and Google AI Studio, with additional availability in Android Studio and Google Antigravity. Enterprise access includes Gemini Enterprise products. Google also says Gemini Spark uses 3.7 Flash for Pro and Ultra subscribers in supported countries.
Pricing requires a dated record
Google lists introductory API pricing of $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. It states that prices rise on January 1, 2027 to $1.50 input and $7.50 output per million tokens. Confirm live prices, caching, tool, grounding, media, and regional charges before production.
Safety and evidence limits
Google says the release includes updated safeguards for cyber and chemical, biological, radiological, and nuclear misuse. Provider benchmarks are useful release context but cannot predict reliability in a specific workflow. High-stakes documents, code changes, and agent actions still require source checks, tests, permissions, logs, and human approval.
How to evaluate it
Run the same representative tasks on the current production model and 3.7 Flash. Record correct completion, tool-call success, citation quality, retries, latency, tokens, accepted output, and full cost. Test permission failures and recovery paths as carefully as successful runs.
Bottom line
Gemini 3.7 Flash is a meaningful candidate for scalable coding and agent workloads, but the introductory price creates a future budget change. Adopt it only after regression testing and a cost model that includes the January 2027 rate.
Read the Gemini review, ChatGPT vs. Gemini, and AI subscription comparison guide.
Our editorial read
The strongest choice when the task benefits from Google context, multimodality and connected productivity surfaces.
Best fit
Users already working across Google Search, Drive, Gmail or Workspace who need assistance near those materials.
Where we would hesitate
Do not assume connected data access is enabled for every account, region or plan.
Decision rule
Choose Gemini for Google-native workflows; choose ChatGPT or Claude when their project, coding or writing ecosystem fits better.
Evidence note: This judgment synthesises the linked official documentation, release notes and product context. It is not a claim that AI Tools Manual independently benchmarked every tool.
Official source check
Official sources used for product facts: gemini.google.com, support.google.com. These links support named features, access, plan or policy statements; they are not independent benchmark results.
The links above are included because this page names or compares these products. They support the documented product facts only; the editorial judgment, workflow advice and limitations remain our own and are not presented as independent test results.
August 25, 2026: intelligent dictation on macOS
Google says the Gemini app for macOS now supports intelligent dictation from across the desktop. Users can speak into Gemini while working in another app; the feature transcribes the speech, removes filler words, handles corrections, and inserts polished, formatted text at the cursor. That makes it relevant for drafting email, notes, and structured copy without switching away from the current window.
This update changes the input workflow, not the evidence available for transcription accuracy in every environment. Before relying on it for sensitive or high-stakes material, check the macOS app version, account and region rollout, microphone permissions, retention controls, and whether audio or transcripts are used for model improvement. We have not independently benchmarked recognition quality or latency.
Primary source: Google: Enable intelligent dictation in Gemini for macOS.
August 26, 2026: Gemini Live and Gemini 3.5 Transcribe
Google announced Gemini 3.5 Transcribe for intelligent real-time and pre-recorded speech workflows. The documentation describes separate Live API and Interactions API paths, formatted transcription, custom vocabulary, word-level timestamps, speaker attribution, and support for more than 85 languages. Google says the model is in public preview for developers and enterprises, with consumer access in selected Gemini surfaces.
A related Gemini Live announcement adds voice-driven productivity workflows such as Daily Brief, hands-free Gmail actions, Personal Intelligence, and Spark tasks that can run across Google apps. These are powerful claims with account, plan, and rollout dependencies. Confirm the exact endpoint, preview terms, microphone and connected-app permissions, retention settings, and human approval for actions that change mail, files, or calendars. Google’s benchmark figures are provider-reported and we have not independently tested transcription accuracy, latency, or agent reliability.
Primary sources: Gemini 3.5 Transcribe and Gemini Live productivity features.
September 2026 update: Gemini 3.8 Flash
Google DeepMind now lists Gemini 3.8 Flash as generally available for agentic coding, knowledge work, multimodal understanding, and high-volume workflows. The official model page lists text, image, video, audio, and PDF inputs, a 1M-token context window, up to 64K output, function calling, search as a tool, and computer use. It is available through the Gemini app, Google AI Studio, the Gemini API, Gemini Enterprise Agent Platform, Gemini AI Mode, and Google Antigravity.
Google’s published tables report higher results than Gemini 3.7 Flash on several internal or partner evaluations, including DeepSWE, Vals Finance Agent, and HLE-Verified. Those figures are vendor-published comparisons, not an independent AI Tools Manual benchmark. The page also notes introductory API pricing through December 31, 2026, so teams must recheck rates after that date.
Gemini 3.8 Flash Cyber
DeepMind separately presents Gemini 3.8 Flash Cyber for trusted defenders through the Fairwind Program. It focuses on vulnerability discovery and automated patching. Access is restricted, and the security claims should be treated as a deployment capability requiring authorization, code review, isolation, and rollback—not as permission to scan systems without approval.
Best fit and limitations
- Best fit: teams building multimodal agents, document-heavy workflows, or coding systems that can use a large context window.
- Not fit: projects that need a fixed price, unrestricted cyber access, or a guarantee that model-generated code is production-safe.
- Check model routing, token limits, caching, data handling, retention, and regional access before migration.
Sources: Gemini 3.8 Flash, Gemini 3.8 Flash Cyber, and the model card. See our Gemini review and Perplexity vs Gemini. No independent test is claimed.
Gemini 3.8 Live adds a real-time voice path
Google’s September announcement introduces Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking as new audio models for more natural voice collaboration, real-time visual context, and background task execution. The standard Live model is positioned for scale and cost efficiency, while Extended Thinking is designed for higher-complexity tasks that need more multi-step reasoning.
The rollout is split by audience. Google says developers can use both models through the Gemini API and Google AI Studio. Gemini 3.8 Live is also rolling out in Search Live, while Extended Thinking is available in Gemini Live and to Google AI Pro and Ultra subscribers in selected Workspace surfaces. Enterprise access is described as private preview or coming soon, so a public announcement should not be read as universal availability.
Best fit and limits
Live is a good fit for voice-first exploration, meetings, visual troubleshooting, and assistants that need to keep a conversation moving while tools work in the background. Extended Thinking is more appropriate when the task can tolerate additional latency in exchange for deeper reasoning. Neither label guarantees reliable transcription, grounding, or action execution for every language, device, or connected tool.
Teams should check region, plan, API quota, latency, audio retention, connected-app permissions, and whether a background action can be interrupted. Voice systems also need a clear confirmation step before sending messages, editing documents, or taking other consequential actions. Google’s page describes capabilities and rollout channels, not an independent accuracy or safety certification.
Our decision rule
Start with a low-risk spoken workflow: ask for a transcript or visual summary, inspect the result, and only then allow a connected action. Choose standard Live for interactive speed; reserve Extended Thinking for tasks where an extra reasoning pass is worth the delay and cost. Compare the Gemini tool profile with the Perplexity vs. Gemini comparison before standardizing on a research workflow.
Evidence disclosure
The primary source is Google’s official Gemini 3.8 Live announcement. AI Tools Manual did not independently test voice latency, visual grounding, background tasks, or plan availability. Verify current rollout, data controls, quotas, regional access, and action permissions before approval.
September 30, 2026: Google announces Gemini 4 Argon
Google announced Gemini 4 Argon as a frontier model for long, multi-step workflows spanning software engineering, legal and finance work, and defensive cybersecurity. The practical news is the staged access model, not general availability: Google says Argon is first rolling out to trusted cyber defenders through its Fairwind Program, with broader access planned for developers, enterprises, and consumers later. The company says it is strengthening safeguards and gathering feedback before that expansion. Readers should not assume that the model is already selectable in the consumer Gemini app or available to every API account.
What Google reports—and what remains unverified
Google says Argon’s output limit is increasing from 64,000 to 1 million tokens and describes strong results on coding, finance, legal, automation, video-understanding, and vulnerability-remediation evaluations. Those figures and rankings are Google-reported results, and some cited evaluations are internal or partner-run. They are useful signals for deciding what to test, not independent evidence that Argon will outperform alternatives on a buyer’s own workflow. The announcement also describes security-sensitive capabilities, including finding and patching vulnerabilities; access is deliberately restricted while safeguards are developed.
Google lists an introductory API price of $2 per million input tokens and $10 per million output tokens, with cached input at a 95% discount from input pricing. It says the later price will be $4 per million input and $20 per million output tokens after the introductory period. Because pricing and access are release-stage facts, verify the live API catalog and terms at the time of purchase; the announcement does not provide a universal date for the end of introductory pricing.
Who should evaluate it now
Argon is relevant now to eligible cyber-defense partners and teams planning an evaluation, especially where long-context analysis or multi-step coding work is important. For other developers, the sensible move is to track availability and prepare a task-specific evaluation rather than build a production dependency around an announced rollout. It is not a fit for buyers who need immediate broad access, a stable public service commitment, or independently verified performance figures. Teams handling security work should also confirm the access program’s conditions and safety controls before submitting sensitive code or incident data.
Our decision rule
When access opens, compare Argon against the model you currently use on a fixed set of representative tasks. Record quality, failure severity, latency, token consumption, human review time, and total cost—not just benchmark scores or maximum context size. Keep security-critical changes behind code review and staged deployment even if a model proposes a patch. For a broader side-by-side framework, see ChatGPT vs. Gemini and our AI-agent evaluation guide.
Evidence disclosure — verify before approval
The primary source is Google’s Gemini 4 Argon announcement. It supports the model description, staged rollout, token limit, announced pricing, safeguards, and vendor-reported benchmark results. AI Tools Manual did not independently run Argon or validate the benchmarks, prices, rollout schedule, or security efficacy. Price and security claims require careful rechecking before publication.
October 2026: Guided Vision and a broader Connected Apps rollout
Google has announced two Gemini changes that matter for different jobs. On October 1, it launched Guided Vision in Gemini Live for compatible Android devices. On September 23, Google described a wider rollout of Connected Apps that brings additional productivity, creative, and lifestyle services into Gemini. These are separate features: one uses a shared camera for real-time visual assistance; the other connects selected services to Gemini conversations.
Guided Vision: conversational camera assistance
With Guided Vision enabled, a user can share the camera during a Gemini Live session, ask follow-up questions about nearby objects or text, and receive spoken cues to adjust the camera framing. Google says the feature was developed with blind and low-vision communities and Aira, and lists examples such as reading small print or locating an object. Google documents access through the Gemini app settings or Android accessibility shortcuts, including TalkBack integration. Availability requires Android 9 or later and is limited to regions and languages where Gemini Live is supported.
Important boundary: Google explicitly says Guided Vision can make mistakes and is not intended for navigation, safe-travel guidance, obstacle detection, or replacing a white cane or other established mobility aid. Treat it as an assistive description tool, not a safety system; independently verify consequential details such as labels, controls, or instructions.
Connected Apps: more work inside Gemini
Google’s September 23 announcement adds services across categories: Airtable, Linear, monday.com, PandaDoc, Wispr AI, and Zoho for productivity; Adobe, Picsart, Squarespace, and Webflow for creative work; and apartments.com, Experian, Peloton, and SeatGeek for lifestyle tasks. Google says users can connect services in Gemini settings or invoke them with an @ mention or a direct request. The rollout means account and feature availability may differ while it expands.
Who benefits, and what to verify first
Guided Vision is most relevant to blind and low-vision users who want a conversational way to inspect nearby visual information, and may also help with low-risk tasks such as reading fine print. Connected Apps are more relevant to people who already manage projects, create assets, or plan activities in the named services. Neither announcement establishes that every account has every capability, nor does the feature description itself settle questions about connector permissions or data handling.
Before enabling a connected service, review the account-selection and permission prompts, choose only the services needed for the task, and avoid sending sensitive records until the available controls and data practices are clear. For camera assistance, use the feature in a safe stationary setting first and do not rely on it to cross roads, avoid hazards, or make medical decisions. Our AI-tool privacy review guide offers a checklist for connected services; our ChatGPT vs. Gemini comparison helps put Gemini’s broader assistant workflow in context.
Our assessment and evidence limits
Guided Vision is a meaningful accessibility-oriented use of live multimodal interaction, but its safety disclaimer should shape the use case: it can support description and exploration, not independent mobility. Connected Apps may reduce context switching, yet convenience depends on the quality of permissions, account boundaries, and rollout coverage. Google’s announcements describe intended capabilities; AI Tools Manual did not test the features, verify a specific user’s eligibility, or audit connector data flows. Confirm current availability and prompts in the product before relying on either workflow.
Primary sources: Google’s Guided Vision in Gemini Live announcement and Connected Apps rollout announcement. Next review: November 3, 2026, or sooner if device eligibility, regional access, safety guidance, or connector permissions change.
October 13, 2026: Gemini 4 Argon reaches Pro and Ultra app subscribers
Google’s October 13 availability note says Gemini 4 Argon is available in the Gemini app to Google AI Pro and Ultra subscribers. It also describes switching from chat to background tasks, progress updates on longer work, connections to more than 40 apps, and access to Google Antigravity’s coding interface from the Gemini macOS app. This is a later availability update to the September 30 announcement, which initially described a trusted cyber-defender rollout. Read Google’s availability note alongside the original Argon announcement.
The distinction matters for buyers: app access is tied to Pro/Ultra, while developer, enterprise and security-program access have separate routes and terms. A feature shown in the consumer app does not imply identical API access, quotas, data controls or regional availability. The new task mode can take actions through connected services, so users should inspect app permissions and keep approval steps for consequential actions.
Our assessment: this changes Argon from an announced, restricted model into a usable app option for eligible subscribers, but it is still not a universal free-tier capability. Google’s performance and safety statements remain vendor claims; AI Tools Manual has not run the model or audited its connected-app data flows. Recheck current API pricing, token limits, country availability and plan entitlements before purchase, especially where the page already discusses time-sensitive API pricing. Compare ChatGPT and Claude for other assistant workflows. Next review: November 16, 2026.
Human review required: reconcile the current Gemini app, API and enterprise availability; verify Pro/Ultra eligibility, regions, connected-app permissions, task controls, current API pricing and the status of the restricted cyber-defense program.
October 28, 2026: Gemini 4 Carbon enters a restricted cyber-defender rollout
Google announced Gemini 4 Carbon as a frontier model aimed at long-horizon coding and cybersecurity tasks. The company says access is initially limited to a subset of cyber defenders and trusted partners through its Fairwind process, with broader developer, business, and consumer availability planned later. The announcement therefore does not mean Carbon is currently available to every Gemini user or API developer. Google also describes Genie 4 as a separate world-model research project announced October 27; it is not the same product or general Gemini assistant model.
Google reports results on security and coding benchmarks and says Carbon is designed to improve secure code generation, reduce unsupported claims, and resist prompt injection. Those are company-reported evaluations and should not be read as an independent security certification or proof that outputs can be deployed without review. The model is explicitly being rolled out in stages while the company gathers feedback on safeguards, which is an important limitation for buyers considering near-term adoption.
Pricing, eligibility, and a decision rule
Google lists introductory API pricing of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at a 95% reduction from the input rate. The announcement says that after the introductory period expires on November 30, 2026, the public-access model price will become $4 per million input tokens and $20 per million output tokens. These are time-sensitive billing claims: verify the current API pricing page and effective date before budgeting, and do not assume that a future Gemini app subscription will include API access at these rates.
Carbon is most relevant to organizations able to qualify for restricted access and evaluate long-running coding or cyber-defense workflows with their own controls. It is not currently a default choice for consumers or teams unable to access Fairwind. Our assessment is to treat Carbon as a high-capability early-access option requiring security review, not as a replacement justified by benchmark headlines. Use non-sensitive evaluation tasks, keep human approval on code changes, and verify model availability before rewriting a workflow. Compare the existing ChatGPT profile and Qwen review for alternatives. Evidence note: All launch, access, benchmark, and price details are attributed to Google’s announcements; AI Tools Manual has not independently tested Carbon. Next review: November 12, 2026.