Google DeepMind Releases Three New Gemini Models, but Not 3.5 Pro

Google DeepMind released three new Gemini models on July 21, but the long-awaited Gemini 3.5 Pro was not among them. Instead, Google introduced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a restricted cybersecurity model called Gemini 3.5 Flash Cyber.

The release gives developers faster and more specialized options for running AI agents. It also creates an awkward contrast. Google is improving the models designed for frequent, lower-cost tasks while its next flagship remains in partner testing.

That omission matters because Google previously indicated that Gemini 3.5 Pro would follow Gemini 3.5 Flash. A model delay report published before this launch said the flagship was months behind schedule, with coding performance among the reported concerns.

OpenAI and Anthropic have continued advancing their own coding and agent models during that delay. Google now has a stronger deployment portfolio, but it still lacks the flagship release needed for a direct response at the top of the market.

The central question is therefore larger than which benchmark moved by several points. Google DeepMind releases three new Gemini models but not 3.5 Pro because practical efficiency is becoming a competitive advantage. Yet efficiency cannot completely replace frontier performance, especially for developers choosing a model for difficult, autonomous work.

Google DeepMind Releases Three New Gemini Models but Not 3.5 Pro

Google’s release strengthens the Flash family while leaving its most visible product gap unresolved.

The new lineup divides production workloads across three distinct models. Gemini 3.6 Flash targets complex coding, multimodal analysis, and multi-step agents. Gemini 3.5 Flash-Lite focuses on high-volume tasks, while Gemini 3.5 Flash Cyber handles vulnerability research inside Google’s CodeMender security agent.

According to Google’s Gemini model announcement, Gemini 3.6 Flash uses 17 percent fewer output tokens than Gemini 3.5 Flash on the Artificial Analysis Index. Output tokens are the text and structured data generated by a model after receiving a request.

Token efficiency matters because an agent rarely makes only one model call. It may inspect files, plan changes, run tools, evaluate results, and repeat failed steps. A reduction across each interaction can compound during a long workflow.

Google says Gemini 3.6 Flash also requires fewer reasoning steps, tool calls, and conversational turns for multi-stage tasks. That claim addresses a common problem with AI agents: a model can appear inexpensive per request but become costly when it repeatedly revises its plan.

The company reports several benchmark improvements over Gemini 3.5 Flash. On DeepSWE, which evaluates software engineering work, the reported score rose from 37 percent to 49 percent. MLE Bench increased from 49.7 percent to 63.9 percent, while OSWorld-Verified moved from 78.4 percent to 83 percent.

Those figures come from Google’s release materials and should be treated as vendor-reported results. Benchmarks help compare controlled tasks, but they do not fully measure reliability inside a company’s codebase, document collection, or internal approval process.

The new model is already generally available through the Gemini API. It also appears in Google AI Studio, Android Studio, the Gemini app, and several enterprise products. Google has made it the default model for its Antigravity agent within Gemini Managed Agents.

Gemini 3.5 Flash-Lite fills a different role. Google describes it as the fastest model in the Gemini 3.5 family, reaching 350 output tokens per second in the Artificial Analysis Index. It is intended for document extraction, structured data processing, agentic search, and fleets of lightweight subagents.

Google says Flash-Lite improved from 31 percent to 54 percent on Terminal-Bench 2.1. It also reports gains on long-context evaluation and real-world task execution. The model can use different thinking levels, letting developers choose between minimal reasoning and more involved multi-step processing.

Both general-purpose models support a context window of more than one million input tokens. A context window is the amount of material a model can consider during one interaction. Their maximum output is 65,536 tokens, according to the Gemini API specifications.

These capabilities make the launch substantial even without Gemini 3.5 Pro. Google has shipped two production-ready models and one specialized security system. The missing flagship, however, changes how the announcement will be interpreted.

Google says Gemini 3.5 Pro is testing with partners and will become broadly available when ready. It also disclosed that pretraining has begun for Gemini 4. Neither statement provides a release date for the delayed model.

That leaves developers with a clear product update and an unresolved strategic question. Google can now handle more routine agent tasks efficiently, but its strongest next-generation model remains outside general availability.

The AI Model Race Is Moving From Answers to Completed Work

Gemini 3.6 Flash reflects a market where the cost of finishing a workflow matters more than the cost of producing one response.

Earlier model launches often centered on broad knowledge tests, math scores, or chatbot comparisons. Production buyers now pay closer attention to whether an agent completes a task without spiraling through unnecessary calls.

This change favors models that balance intelligence, latency, and predictable behavior. A slightly weaker model can be more useful if it finishes common work quickly and avoids repeated corrections. A stronger model can become impractical when every workflow requires long reasoning traces and expensive retries.

Gemini 3.6 Flash is designed around that distinction. Google is not only claiming higher benchmark scores. It is emphasizing fewer unwanted code edits, shorter execution loops, and lower token consumption.

Those improvements target problems developers can observe directly. An agent that changes unrelated files creates review work. An agent that keeps calling tools increases latency. An agent that produces lengthy explanations can consume resources without advancing the task.

Google’s documentation also says the model more often runs diagnostic scripts before editing code. That behavior can improve accuracy during difficult technical work because the model gathers evidence before choosing a fix.

There is a tradeoff. The same documentation warns that upfront inspection can add unnecessary steps to simple frontend tasks. Human evaluators also preferred earlier models for some visual styling work, despite Gemini 3.6 Flash producing more functional code.

This qualification is important. Google DeepMind releases three new Gemini models but not 3.5 Pro at a moment when efficiency claims need workload-level testing. A model can improve overall while becoming less suitable for particular design, writing, or low-complexity tasks.

Gemini 3.5 Flash-Lite extends the efficiency strategy. It can act as a worker model beneath a more capable coordinator. One agent might plan a job while many Flash-Lite instances extract records, inspect documents, or test candidate outputs.

That architecture resembles a workplace team more than a conventional chatbot. The coordinator handles ambiguous decisions. Smaller workers perform bounded assignments in parallel. The system then combines their results.

Google illustrated this approach by pairing Gemini 3.6 Flash with Flash-Lite subagents that generated multiple web design concepts. It also highlighted receipt processing, large-scale product feature extraction, and document analysis as possible workloads.

The commercial appeal is straightforward. Companies do not need their most capable model for every step. They need a system that assigns the appropriate level of intelligence to each step without losing control of the overall process.

OpenAI has adopted a similar portfolio strategy. Its GPT-5.6 family includes models aimed at different balances of intelligence and efficiency. Anthropic also separates flagship reasoning from faster models suited to frequent coding and workplace tasks.

That makes Google’s Flash expansion competitive rather than unusual. The difference is timing. Google is broadening its efficient model portfolio while competitors already have newer flagship options in general circulation.

The pressure falls most heavily on Google’s enterprise AI position. Buyers comparing platforms want strong coding, computer use, multimodal understanding, security controls, and predictable operating costs. They also want confidence that the vendor’s flagship roadmap will remain current.

Gemini 3.6 Flash addresses the operating side of that equation. The missing Gemini 3.5 Pro leaves the roadmap side less settled.

The Missing 3.5 Pro Is the Real Story

Google’s three-model launch shows execution in production AI, but it does not erase the reported delay of its flagship.

Gemini 3.5 Pro was expected to represent Google’s strongest general-purpose model in the 3.5 generation. It would compete for tasks where deeper reasoning and coding ability matter more than raw speed.

A Bloomberg report said Google was months behind schedule because it was trying to improve the model, particularly its coding capabilities. The report cited people familiar with the matter, not a public technical disclosure from Google.

Google has now confirmed only that Gemini 3.5 Pro is being tested with partners. Its announcement says the company will release the model broadly when it is ready. That wording avoids a revised date and gives Google room to continue development.

The delay becomes more significant because coding has become a central test of frontier models. Coding agents must understand large repositories, plan changes, operate tools, recover from errors, and preserve existing behavior. These requirements expose weaknesses that a short benchmark prompt may hide.

A model can write a plausible function yet struggle to complete a repository-wide migration. It can identify an error but edit the wrong file. It can generate valid code while failing to test the surrounding system.

Gemini 3.6 Flash improves several parts of this workflow, according to Google. Its gains do not necessarily reveal the state of Gemini 3.5 Pro. The two models serve different performance targets and may use different training or post-training approaches.

Still, the Flash release gives Google a practical way to keep developers engaged. Teams can test the new API, update agent systems, and evaluate token efficiency without waiting for the flagship.

It also gives Google additional feedback about real agent behavior. Patterns from Gemini 3.6 Flash deployments can inform future models, especially where tools, code changes, and long interactions create failures.

The risk is that developers make durable platform choices before Gemini 3.5 Pro arrives. Agent systems require more than changing a model identifier. Teams build evaluation suites, permission boundaries, monitoring, prompts, and recovery logic around a provider’s behavior.

Once those systems work, switching can become difficult. A delayed flagship therefore affects more than launch-day attention. It can influence which platform developers choose for their next production workflow.

OpenAI’s newer flagship family increases that pressure. Anthropic’s coding products provide another established alternative. Open-weight models add a separate challenge for organizations that prefer to run models within their own infrastructure.

Google has several advantages. It controls widely used developer tools, cloud infrastructure, Android, Search, and a large consumer AI product. Gemini can reach users through distribution channels that most model providers cannot match.

Distribution, however, does not guarantee preference among developers building advanced agents. Those teams often select models task by task. They can use one provider for routine document work and another for difficult coding.

This is why the omission cannot be dismissed as a naming detail. Google DeepMind releases three new Gemini models but not 3.5 Pro while the market is rewarding both efficient workers and highly capable coordinators.

Gemini 3.6 Flash strengthens Google’s worker and midrange position. Gemini 3.5 Pro still needs to prove that Google can lead the most difficult layer of the same system.

Gemini 3.5 Flash Cyber Makes Specialization the Bigger Bet

The cybersecurity model shows that Google sees specialized agent systems as another path to frontier-level results.

Gemini 3.5 Flash Cyber is built on Gemini 3.5 Flash and fine-tuned to find, validate, and patch software vulnerabilities. It will initially operate through CodeMender, Google DeepMind’s security agent.

Unlike Gemini 3.6 Flash and Flash-Lite, the cyber model will not receive a broad public release. Google plans a limited pilot for governments and trusted partners because vulnerability discovery has clear defensive and offensive uses.

That deployment decision reflects a difficult reality. A model that helps maintainers find flaws can also help attackers locate weaknesses before patches become available. Wider access increases the number of potential defenders and the number of potential abusers.

Google’s cybersecurity model details describe CodeMender calling the specialized model multiple times for one investigation. Separate agents explore code paths, validate possible vulnerabilities, and combine their work into a final report.

This approach turns efficiency into a technical mechanism, not simply a budget benefit. Vulnerability research involves a large search space. A faster model can inspect more paths and test more hypotheses within the same operational window.

On the CyberGym benchmark, Google configured CodeMender to call Gemini 3.5 Flash Cyber up to five times for one final report. The company says the resulting agent achieved competitive performance against much larger models.

Google included an important limitation. Competitor results on that benchmark came from providers’ self-reported scores. Differences in tools, prompts, and agent scaffolding can affect comparisons, so the result should not be read as a controlled model-only ranking.

The company also tested the model on Chrome’s production commit-scanning pipeline. Those vulnerabilities had not been publicly disclosed, reducing the risk that the test material had appeared in training data.

On a separate evaluation involving the V8 JavaScript engine, Google says Flash Cyber found 55 confirmed issues. Mainline Gemini 3.5 Flash found 47, while Claude Opus 4.6 found 36. The specialized model identified 10 issues missed by both comparison models.

Those are company-reported findings, and independent researchers have not yet reproduced the full evaluation. Google also noted that newer competitor models refused some tasks because of their safety controls, which complicates direct capability comparisons.

Even with those caveats, Flash Cyber illustrates a meaningful shift. The strongest result for a narrow task may come from combining a smaller specialized model with repeated calls and a carefully designed agent.

That idea challenges the assumption that every difficult problem requires the largest available general model. A specialized model can spend its capacity on one domain, while the surrounding system supplies tools, coordination, and verification.

The restricted pilot also reveals the tradeoff behind specialization. Google can give selected defenders a head start, but most developers cannot test the model. External researchers will have limited access to assess false positives, missed vulnerabilities, or misuse controls.

Governments and trusted partners will therefore shape the first evidence about its practical value. Their results will matter more than benchmark charts, especially if CodeMender finds and helps repair previously unknown flaws.

The model’s success will depend on the entire security workflow. Finding a suspicious code path is only the beginning. A useful system must confirm the vulnerability, propose a safe patch, avoid regressions, and communicate evidence to human maintainers.

Flash Cyber is therefore not just a third model attached to the announcement. It is Google’s clearest argument that model specialization and agent orchestration can compensate for the limits of a general-purpose model.

Google’s Benchmark Gains Still Need Production Proof

The release offers credible signs of progress, but Google’s measurements cannot answer every question developers face.

Benchmark improvements provide a useful starting point. DeepSWE tests coding behavior, OSWorld evaluates computer use, and long-context tests measure whether a model can recover information from large inputs.

Each benchmark isolates a slice of performance. Production agents combine many slices at once. They must follow instructions, maintain state, choose tools, handle permissions, and stop when the job is complete.

Small failure rates can accumulate during these workflows. An agent that succeeds on most individual steps can still fail frequently when a task requires dozens of dependent actions.

Token reduction also needs careful interpretation. Fewer output tokens can mean better focus and less wasted reasoning. It can also mean incomplete explanations, skipped checks, or shorter reports.

Google’s own documentation identifies workload-specific weaknesses. Gemini 3.6 Flash may perform extra diagnostic work on simple frontend requests. Human evaluators preferred earlier models for some visual layout and styling tasks.

These disclosures make the release more useful because they help developers design realistic evaluations. They also show why a single overall ranking cannot determine the right model.

Teams should test complete tasks drawn from their actual environment. For coding, that means repository-level changes with tests and human review. For document analysis, it means real file formats, charts, incomplete records, and conflicting evidence.

For computer use, evaluation should include recovery from interrupted pages, changed interfaces, and ambiguous controls. A model that performs well on a stable benchmark may struggle when a website changes without warning.

Organizations also need to measure unnecessary actions. Every extra tool call can create latency or risk. An agent with permission to edit files, send messages, or access customer information needs stricter boundaries than a chatbot.

The new API introduces migration work as well. Google says developers must remove several deprecated sampling parameters and prefilled model turns. Some applications will need to change how they maintain multi-turn conversations.

These changes are manageable, but they complicate immediate comparisons. A team that moves an existing application to Gemini 3.6 Flash may be evaluating model quality and API behavior at the same time.

The 17 percent token reduction deserves similar caution. It comes from an external index cited by Google, but the benefit will vary across prompts, tools, and thinking settings. A concise coding agent may show different savings from a detailed research assistant.

Gemini 3.5 Flash-Lite presents another evaluation challenge. Its high throughput is attractive for document processing, yet speed alone does not establish extraction accuracy. Errors repeated across thousands of documents can become expensive to detect later.

Businesses should measure field-level accuracy, citation quality, and exception rates. They should also track whether a larger model must repeatedly correct Flash-Lite’s work. A fast worker loses its advantage when a coordinator must redo every difficult case.

The same principle applies to Flash Cyber. Multiple model calls can expand the search space, but they can also generate duplicate or low-quality findings. Human security teams need evidence that the combined report prioritizes real vulnerabilities.

None of these concerns invalidate Google’s claims. They define the verification work that follows a model launch.

The missing Gemini 3.5 Pro adds another uncertainty. Developers do not know whether Google’s next flagship will preserve the efficiency improvements, use compatible interfaces, or require another round of system tuning.

Google DeepMind releases three new Gemini models but not 3.5 Pro, giving teams useful products now and an incomplete roadmap for later. The strongest response is neither immediate adoption nor automatic dismissal. It is controlled testing against the workloads that matter.

Three Signals Will Decide Whether Google’s Strategy Works

The next phase will be determined by the 3.5 Pro release, real agent economics, and evidence from the restricted cyber pilot.

The first signal is the availability and quality of Gemini 3.5 Pro. Google says partner testing is underway, but it has not committed to a public date.

A broad release within the next few months, supported by strong coding and agent evaluations, would reduce the importance of the current delay. It would let Google pair an efficient worker portfolio with a current flagship coordinator.

Another delay would strengthen the opposite interpretation. It would suggest that Google can optimize existing Flash models more predictably than it can deliver the next performance jump at the top of the range.

The second signal is production behavior from Gemini 3.6 Flash and Flash-Lite. Developers should watch task completion rates, tool-call counts, latency, unwanted edits, and correction requirements.

If teams complete more work with fewer retries, Google’s efficiency argument will become much stronger. The 17 percent token reduction would then represent a broader improvement in agent economics, not merely shorter output.

If applications require more supervision or repeated calls, benchmark and token gains will look less decisive. The crucial unit is the completed, verified task.

This distinction will shape purchasing decisions across coding, research, document processing, and customer operations. Enterprise buyers increasingly evaluate the whole workflow rather than the model response in isolation.

The third signal is evidence from the Gemini 3.5 Flash Cyber pilot. Google has positioned limited access as a safety measure that gives defenders an early advantage.

That claim will gain support if trusted users identify meaningful vulnerabilities, produce reliable patches, and disclose results through responsible security processes. Evidence from real codebases would validate the specialized, multi-agent design.

The case weakens if results remain limited to vendor benchmarks. It would also face pressure if false positives overwhelm reviewers or if access restrictions prevent credible external assessment.

Competitor responses will shape all three signals. OpenAI and Anthropic can reduce Google’s efficiency advantage by improving their smaller models. Open-weight providers can pressure every closed platform by offering greater deployment control.

Google’s distribution remains a major asset. The new models can reach developers through familiar tools and consumers through Gemini and Search. Broad availability gives Google a large base for feedback and iteration.

Yet distribution will not settle the flagship question. Developers building difficult agents often compare several providers and route tasks to different models. They will choose Gemini when its complete system performs better for their workload.

For knowledge workers, the release is also a reminder that model choice is only one part of a reliable AI workflow. The information supplied to an agent, its permissions, and its ability to cite evidence remain equally important.

Tools that organize work context through knowledge blending can help users maintain that evidence layer across documents and notes. The underlying model may change, but the quality of accessible context still shapes the result.

Google DeepMind releases three new Gemini models but not 3.5 Pro because the company has useful improvements ready before its flagship. That is a rational product decision, but it is not a complete strategic answer.

The next few months will show whether Google has built a durable efficiency lead or bought time while finishing Gemini 3.5 Pro. Developers should test the models now, record task-level results, and compare them again when the missing flagship finally arrives.