Grok vs Kimi vs Gemini vs Codex: which fits your work?

Find the answer you need.

First, separate four layers

Model

A model performs the reasoning and generation. Examples in this guide include Grok 4.7, Kimi K3, Kimi K2.7 Code, Gemini 3.8 Flash and the GPT models available inside Codex.

Assistant

An assistant is the user product around a model. It adds an account, conversation interface, files, memory, search and connected apps. The Grok, Kimi and Gemini consumer or business apps belong here.

Coding agent

A coding agent adds a working environment for software. It can inspect a repository, edit files, run a shell, execute tests, use tools and return a diff or pull request. Kimi Code and Codex belong here, though each can use different models and deployment modes.

Platform

A platform gives a business APIs, tool contracts, identity, workspaces, logging, spending controls, deployment options and agent management. SpaceXAI, formerly xAI and still using the x.ai developer domain, provides the Grok API platform. The Kimi API, Google's Gemini API and enterprise products, and OpenAI's separate Agents API also belong here.

Use this test whenever someone says one product is better: which exact model, inside which assistant or agent, with which tools, plan, permissions, source data, region and budget? Without those details, the claim is too vague to buy or build around.

What the four names mean on 7 October 2026

Grok 4.7 facts come from SpaceXAI's current model page. Current information is not inherent in the model; web or X search must be enabled for a task that needs it. (SpaceXAI, Grok 4.7) (SpaceXAI model list)

Moonshot calls Kimi K3 its flagship model and documents K2.7 Code as the coding-focused route. Kimi Code is the runtime around the model and can also be configured with providers other than Moonshot. Moonshot's K3 launch material says its own evaluation placed K3 behind Claude Fable 5 and GPT-5.6 Sol overall. Treat that as a vendor-run, mixed-source comparison, not a controlled cross-vendor benchmark. (Moonshot AI, Kimi K3) (Moonshot AI, Kimi K2.7 Code) (Kimi Code, getting started)

Google documents Gemini 3.8 Flash as a stable model for long-horizon software engineering, autonomous agents and enterprise workflows. It accepts several media types but returns text; separate Gemini-family models handle image generation. Google's managed agent product remains in preview. (Google, Gemini models) (Google, Gemini 3.8 Flash) (Google, managed agents)

OpenAI documents Codex as a coding-agent product, while its model guide recommends GPT-6.1 Sol for complex coding and agent work when available, Astra for the hardest end-to-end jobs and Luna for focused high-volume tasks. The model does the reasoning; Codex supplies a working environment and software-delivery flow. The OpenAI Agents API is a separate product surface. (OpenAI model guide) (OpenAI Codex CLI) (OpenAI Codex IDE) (OpenAI Codex cloud environments) (OpenAI Agents API)

At-a-glance workflow fit

This is a shortlist, not a rank. It describes where documented capabilities create a sensible first test. A real pilot can still produce a different winner.

This table omits prices because API tokens, search fees, consumer subscriptions, business seats and usage credits are not comparable units. Compare total cost per accepted result in a real pilot.

Grok: strong when live web and X are part of the job

Good fit

Grok deserves a first test when a workflow depends on current web discussion, X search, document search, code execution and a large context window. SpaceXAI's file system can turn a request into an agent-style document-search flow and combine analysis with code execution. That combination suits fast-moving market research, social listening, document packs and research that needs calculations. (SpaceXAI Files)

Where it can fail the fit test

Search must be configured. A model label does not guarantee current evidence, and access to X does not make a source reliable. Research still needs a source hierarchy, dates and a reviewer who can reject unsupported synthesis.

Data settings also change the available features. SpaceXAI says API requests are retained for 30 days by default. SpaceXAI documents team-level zero data retention, but self-service availability can vary by environment or enterprise contract. Enabling it disables stateful Responses, Files, Collections and Batch. Its United States endpoint guarantee excludes server-side tools and files. Here, endpoint means the exact API address and region through which data is processed. A buyer has to inspect that endpoint and the enabled feature set instead of attaching one privacy statement to the whole Grok brand. (SpaceXAI model documentation) (SpaceXAI security FAQ) (SpaceXAI regions)

DLVX view

We would use Grok when X is a material source or when its live research toolset clearly reduces collection work. We would not use it as the sole judge of the evidence it found. The useful output is a dated source pack, not an uncited opinion about what the internet thinks.

Kimi: strong for long context, coding sessions and open models

Good fit

Kimi K3 belongs in tests involving very large source packs, vision and long-horizon knowledge or coding work. Kimi K2.7 Code narrows the focus to software engineering, while Kimi Code provides repository access, shell tools, web retrieval, skills, MCP and subagent support. Moonshot explicitly recommends different models for coding and general writing or analysis, showing that routing is needed even inside one vendor. (Kimi K3) (Kimi K2.7 Code)

The open-model route also matters. A capable engineering team can evaluate self-hosting K3 or K2.7 Code weights when control or customisation justifies the infrastructure. That is a different product decision from using the managed Kimi API. (Kimi K3) (Kimi K2.7 Code)

Where it can fail the fit test

One million tokens measure capacity, not attention or truth. A long source pack still needs document boundaries, retrieval checks, citations and tests. The managed Kimi API is cloud-only, and Kimi's international platform is separate from mainland China. Moonshot does not promise equal overseas stability in every region. (Kimi API troubleshooting)

Open weights are not free operations. Serving, monitoring, updates, inference hardware, access control and incident handling move cost from the API invoice into the engineering budget.

DLVX view

We use Kimi-style routes where long context or sustained implementation makes them efficient, then verify the result through tests or another reviewer. We do not infer current model superiority from the older internal route name. Product updates can change the decision, which is why the task pack and acceptance test remain portable.

Gemini: strong when the business already runs through Google

Good fit

Gemini deserves the first evaluation when source material and daily work already live in Google Workspace. Gmail, Docs, Sheets, Slides, Drive, Chat and Meet access can remove the export-and-upload step that breaks many otherwise sensible AI workflows. Google also offers Search grounding, Maps, URL context, file search, code execution, function calling and structured outputs through developer products. (Google Workspace with Gemini) (Google Gemini tools)

Gemini 3.8 Flash also makes a strong shortlist for applications that need text, image, video, audio and PDF inputs through one stable model route. That breadth can reduce pre-processing and provider count.

Where it can fail the fit test

Native fit is not a universal quality win. Google documents feature-specific data handling: Search grounding stores prompts, supplied context and generated output for 30 days, and that storage cannot be disabled. Workspace, the Gemini Developer API, Vertex AI, Gemini Enterprise and managed agents require separate control checks. Managed agents remain preview, and residency-constrained deployments can expose fewer models or features than the global endpoint. (Google Gemini zero data retention guide) (Google Gemini Enterprise locations)

A Google Workspace business may still get a better coding result from Codex or Kimi Code. Integration convenience and model performance are separate dimensions.

DLVX view

We would start by testing Gemini when it can work inside the client's existing Google permissions and data, because that can reduce new copies and new access paths. We keep repository work and independent review open to other systems. Fewer exports are valuable, but not if the chosen surface fails the data or acceptance gate.

Codex: strong for repository-level software delivery

Good fit

Codex is the clearest first test when the required output is a working repository change: implementation, refactoring, test repair, code review, migration or repeatable developer automation. Codex is available through the ChatGPT desktop app, CLI, IDE extension and Codex Cloud. Separately, the OpenAI Agents API exposes an OpenAI-managed Codex harness through an API. The Agents API manages sessions, orchestration, compaction and recovery; its availability and data controls do not automatically apply to local Codex or Codex Cloud. (OpenAI Codex CLI) (OpenAI Codex IDE) (OpenAI Codex cloud environments) (OpenAI Agents API)

Its real advantage appears when the team already has a clean repository, written acceptance criteria and tests. The agent can inspect the actual system, change files and prove the result instead of returning a code sample detached from the build.

Where it can fail the fit test

Codex is not one model, so a model comparison must name what ran. Context limits, behaviour and cost depend on that selection and the surface. A local Codex interface can still call remote models or external tools; local agent does not mean local inference.

OpenAI says the Agents API currently offers United States-only data residency and does not support zero data retention. OpenAI's HIPAA configuration guide describes a shared-responsibility model for local Codex: OpenAI handles inference inputs and outputs, while the customer configures workstations, repositories, local retention, MCP, browser and computer use, apps and third parties. This guidance is for Codex deployments handling protected health information, not a blanket retention statement for every Codex surface. (OpenAI Agents API) (OpenAI HIPAA configuration guide)

DLVX view

We use Codex for code-native work and review when repository context, tests and a clear diff matter. It is often the wrong tool for an office workflow that never touches a codebase. Even in software delivery, we judge the accepted change, not the number of generated lines or how confidently the agent explains itself.

Why combinations often beat a single subscription

One provider can cover several stages, but forcing it to cover all of them creates a fragile workflow. Better combinations preserve a portable brief and assign each stage to the best available route.

Gemini plus Codex

Gemini can work with Workspace source material, permissions and office documents. Codex can take the approved brief into a repository, implement it and run tests. The handoff should be a versioned brief with source references and acceptance criteria, not a loose chat summary.

Grok plus Codex or Kimi Code

Grok can collect time-sensitive web and X evidence. A coding agent can turn a frozen, cited brief into software. Keep the raw sources because the web evidence may change after implementation starts.

Kimi plus Gemini

Kimi can process an unusually large mixed source pack. Gemini can deliver the final office workflow inside Google Workspace or Google Cloud. This split helps when source volume and final integration point favour different products.

Kimi Code plus Codex review, or the reverse

A second coding agent can challenge architecture, tests and edge cases. This does not guarantee independence, but it avoids the obvious failure where one run writes the change and approves its own assumptions.

No combination should become permanent by habit. Keep the brief, test and output contract outside any provider, then rerun the comparison after a material model or policy change.

How the client environment changes the answer

Google's availability list documents Thailand, but country availability, model access, billing and residency are different questions. Test from the real client account and endpoint. (Google Gemini API regions)

Data, privacy and governance cannot be reduced to one logo

Do not ask whether a vendor is compliant in the abstract. Name the exact product, plan, endpoint, region, identity path, retention setting, connected tool and contract. Then map every data flow. A legal or security review can assess that concrete system; it cannot approve a brand name.

How DLVX chooses for a client

We use Kimi and Codex in current internal routes. We shortlist Grok or Gemini when their documented capabilities fit the client environment, then test them before adoption. We start with the job, not the launch announcement. For every client implementation, we choose the stack that fits the existing systems, data boundary, geography, team skill, governance, budget and goal, then keep the surrounding business process explicit.

Availability is not adoption. A shortlist is not an evaluation, an evaluation is not occasional operating use, and occasional testing is not regular production. The evidence used for this guide supports current Kimi and Codex routes. It does not establish a current Grok workflow or a broad Gemini business route.

Our routing method follows six principles:

1. Keep the source of truth where it already belongs. Native access through approved identity is usually safer than exporting uncontrolled copies.
2. Route by stage. Research, planning, implementation, review and operation can use different products.
3. Make the model replaceable. Store instructions, source packs, tests and output contracts outside the provider conversation.
4. Give agents the minimum authority needed. Read, draft, write, publish, send, delete, deploy and pay are different permissions.
5. Verify the consequence. Sources prove research, tests prove code, rendered inspection proves interfaces and durable receipts prove side effects.
6. Review product changes. Access, pricing, models, limits and policy move quickly. A route keeps its place only while it continues to pass.

This is why our verdict is not fake neutrality. We do prefer certain systems for certain jobs at a given time. We simply refuse to turn a useful current preference into a permanent universal claim.

The eight-part decision framework

1. Job

Name the accepted output: a cited report, tested pull request, updated spreadsheet, classified queue or working automation. “Use AI” is not a job.

2. Context

Locate the source material, its size, owner, freshness and existing permissions. Do not build a new data copy until the task proves it needs one.

3. Tools

List required search, files, shell, browser, code execution, office apps, MCP or custom functions. A capable model without the necessary approved tool is still the wrong system.

4. Control

Check whether files, network, tools, credentials and consequential actions can be restricted. Decide what always runs, what always asks and what is denied.

5. Data

Define what may leave the environment, where it can be processed, how long it can remain and whether the product may use it for training. Include tool and grounding paths.

6. Economics

Measure model use, search and tool charges, retries, latency, reviewer time, integration work and failure cost. Optimise for accepted result, not a cheap first response.

7. Evidence

Decide how the output will be cited, tested, diffed, replayed, audited and rolled back. If none apply, name the accountable human judgement.

8. Portability

Keep enough of the prompt, source pack, acceptance test and output format outside the provider to change routes without rebuilding everything.

Apply hard gates before scoring. Reject any route that fails region, contract, retention or residency requirements. Reject agents that cannot reach the required source through an approved identity, cannot be bounded for planned side effects, or fail the real pilot at the required latency and cost.

A fair pilot, without benchmark theatre

Run one meaningful client task through the shortlist. Use the same frozen source pack, acceptance criteria, permissions, tool budget and time window. Record the exact product, model, mode, date and configuration. Repeat enough times to expose instability.

Measure:

• Acceptance rate without material human rewrite.
• Time to a verified result.
• Total cost per accepted result.
• Factual, code and permission failures by severity.
• Tool-call success and recovery behaviour.
• Reviewer time and rollback effort.
• Adoption by the people who must use the workflow.
• Portability of the brief, tests and artefacts to a second provider.

Vendor benchmarks can inform the shortlist, but different harnesses, tools and budgets make cross-vendor score tables poor procurement evidence. A paid pilot on the client's real job gives a smaller answer that can actually be used.

A simple routing tree

Is the source of truth mainly Google Workspace? Start by testing Gemini for discovery and office work.

Is the required output a tested repository change? Test Codex and Kimi Code on the same issue and test suite.

Does the task require current web or X evidence? Test Grok, Gemini or Kimi with search enabled and enforce a source policy.

Is very long multimodal context or an open-weight route central? Test Kimi K3 and include the infrastructure cost.

Did every option pass the data, region, permission and audit gates? If not, remove the failing route before quality scoring.

Which option wins the paid pilot? Keep that winner for this workflow only. Recheck after a major product, price or policy change.

Five questions businesses ask about Grok, Kimi, Gemini and Codex

Which is best: Grok, Kimi, Gemini or Codex?

None is best at everything. Grok fits web and X-connected research, Kimi fits long context and open-model routes, Gemini fits multimodal and Google-connected work, and Codex fits software delivery. The best choice depends on the job, systems, data rules, region, budget and current product access.

Is Codex the same as GPT?

No. GPT names OpenAI model families. Codex is a coding agent and working environment that runs supported GPT models. The model supplies reasoning and generation; Codex adds the tools, permissions, files and software-delivery flow available on the selected Codex surface.

Should a Google Workspace business choose Gemini?

Gemini deserves the first test because it works inside Gmail, Docs, Sheets, Slides, Drive, Chat and Meet. It is not an automatic winner for every stage. The same business may use Codex or Kimi Code for implementation and another route for independent review.

Can Kimi run on private infrastructure?

Moonshot describes Kimi K3 and K2.7 Code as open models, so capable teams can evaluate operating the weights themselves. The managed Kimi API is cloud-only and does not provide a standard on-premises deployment. Self-hosted weights, Kimi Code and the managed API are different routes.

How should a business compare AI tools safely?

Use one real task, the same source pack, clear permissions, a fixed acceptance test and a review step. Measure accepted output, time, total cost, errors and reviewer effort. Check data handling and region rules for the exact product surface before sending sensitive data.

Download the workflow routing matrix

Use the AI workflow routing matrix to record the workflow, accepted output, source, required tools, data class, candidate routes, approval points, proof, cost limit and review dates.

Update policy

This guide was checked on 7 October 2026. We review it when a named model changes status, a stable release replaces a preview, a price or context limit changes, a product adds or removes a tool, or a retention, residency or availability policy changes. dateModified should move only when the advice or product facts change materially.

Sources

• SpaceXAI, Grok 4.7, checked 7 October 2026.
• SpaceXAI model list, checked 7 October 2026.
• SpaceXAI Files, checked 7 October 2026.
• SpaceXAI API security, checked 7 October 2026.
• SpaceXAI regional endpoints, checked 7 October 2026.
• xAI joins SpaceX, checked 7 October 2026.
• SpaceXAI company page, checked 7 October 2026.
• Moonshot AI, Kimi K3, checked 7 October 2026.
• Moonshot AI, Kimi K2.7 Code, checked 7 October 2026.
• Kimi API model selection, checked 7 October 2026.
• Kimi API overview, checked 7 October 2026.
• Kimi API troubleshooting, checked 7 October 2026.
• Kimi API data security, checked 7 October 2026.
• Kimi API pricing, checked 7 October 2026.
• Kimi Business, checked 7 October 2026.
• Kimi Code getting started, checked 7 October 2026.
• Kimi Code tools, checked 7 October 2026.
• Google Gemini model catalogue, checked 7 October 2026.
• Google Gemini 3.8 Flash, checked 7 October 2026.
• Google Workspace with Gemini, checked 7 October 2026.
• Google Gemini tools, checked 7 October 2026.
• Google Gemini managed agents, checked 7 October 2026.
• Google Gemini zero data retention, checked 7 October 2026.
• Google Gemini API regions, checked 7 October 2026.
• Google Gemini Enterprise locations, checked 7 October 2026.
• OpenAI model guide, checked 7 October 2026.
• OpenAI Codex CLI, checked 7 October 2026.
• OpenAI Codex IDE, checked 7 October 2026.
• OpenAI Codex cloud environments, checked 7 October 2026.
• OpenAI Agents API, checked 7 October 2026.
• OpenAI HIPAA configuration guide, checked 7 October 2026.
• OpenAI pricing, checked 7 October 2026.

Related DLVX guides and services

• Practical guides
• How to choose AI for a real business job
• Your first useful AI workflow
• Claude models for business
• AI tools available in Thailand
• AI integration services
• Pro Pick Property system case

Need a model-neutral implementation plan?

We can map the job, data boundary, tool access and proof standard, then test the routes that fit your environment. Book an AI workflow consultation.