Google Gemini 3.8 Flash-100: fully reviewed deployed scope · 2026-10-14
01
What this page covers
Google Gemini 3.8 Flash-100: fully reviewed deployed scope · 2026-10-14
This overview is translated for international visitors. Detailed technical and contractual wording follows the Chinese version.
1.20.49.0
Google Gemini 3.8 Flash-100: fully reviewed deployed scope
Native GenerateContent, Chat and Interactions support text, image, audio, video and PDF input, thinking, JSON, functions, streaming and continuation. Codex uses the HCD Responses bridge. Knowledge Studio adds isolated Python (NumPy, SymPy and Matplotlib), URL reading, video-frame inspection and controlled web search. These are HCD bridges; native hosted code, search and AGENTIC video loops are not enabled. Computer use returns proposed actions only. Combined context: 1,048,576 tokens; no disclosed independent input maximum. Default and maximum output: 65,536 tokens including thinking, subject to remaining context. Explicit cache read/delete/expiry and owned HCD files, background responses and Batch preserve original results. Batch uses standard call prices. Per million tokens: input 3,000,000 points, cache hits 300,000 and output including thinking 15,000,000. Explicit cache creation uses the input rate; storage costs 2,000,000 points per million token-hours. Supplier promotional credits have not been received and do not promise customer refunds. Performance 76.6 and value 66.1 are estimates: 12 model-wide public metrics and 29 estimated metrics, with no direct performance measurement for this offering. Expensive limits were checked offline. Account and Key charges settle once from original usage. Supplier metered estimates are not cash invoice costs. Update the connector and start a new conversation. Code, URL, video inspection and search use the installed Knowledge Studio; existing plugin settings and permissions remain in effect.
1.20.52.0
doubao-deepseek-v4-1-flash
Context 1,048,576 tokens; input maximum 1,048,576; default/maximum output 393,216 including thought, within the combined context. Compatibility image limit 1,500; native 3,600; each media item 50 MiB, request 64 MiB. Owned inline files: 5 MiB each, 24-hour retention. Implicit cache is metered. Explicit cache, native hosted search, Flex, Batch and background generation are closed and rejected. HCD MCP search reserves and bills separately. Actual Work client remains unverified. Performance 77.5/value 76.8 are estimates over 41 metrics, 12 direct model sources, no direct offering benchmark. Expensive boundaries checked offline. Connector1.20.52.0; 44 accepted model descriptions. Exact model verified within the described deployment scope.
1.20.51.0
Google Gemini 4 Argon: verified deployed capabilities
Text, image, audio, video and PDF input; three reasoning levels, JSON, function tools and streaming. Codex tools and conversation continuation are verified. HCD MCP provides isolated Python, URL reading, video-frame inspection and separately billed controlled search. Work client verification is pending; built-in web_search is not open and explicit requests are rejected. Combined context is 1,048,576 tokens; the independent input limit is not disclosed. Native GenerateContent defaults to and permits 1,048,576 total output tokens through manual owner-bound continuationToken requests; each segment permits 32,768 non-streaming or 65,536 streaming tokens. Responses, Chat and Interactions default and maximum per-call outputs are 32,768 non-streaming / 65,536 streaming, including reasoning within combined context. Explicit cache is unsupported; implicit cache is metered. Owned HCD files, background responses and Batch retain lifecycle controls; Batch uses standard call prices. Per million tokens: 8,000,000 input points, 400,000 cache-hit points and 40,000,000 output/reasoning points. Full input, output and tool liabilities are reserved against both account and Key before submission; each explicit continuation is separately reserved and settled. Performance 77.2 and value 62.5 are estimates: 6 model-wide public metrics plus 35 estimates in a fixed 41-metric panel, with no direct offering performance measurement. Expensive boundaries were checked offline. Supplier promotional credits are not treated as received. Account and Key settle separately once from original usage; supplier estimates are not invoice cash costs.
HCD summary views
Gemini 3.8 Flash-100 · HCD summary views
Gemini 3.8 Flash-100 supports HCD summary views: auto keeps the provider automatic summary; detailed shows the full available provider summary; concise explicitly shows an HCD bounded extract and truncation status; none hides the public summary. These views do not control provider summary length or add model calls. Thinking tokens retain original receipt billing. Actual remote Work roundtrip verification is pending.
HCD controlled Work search
Gemini 3.8 Flash-100 · Work Responses
Work Responses live web search uses an HCD controlled bridge. In auto mode the model decides whether to search, at most once per response; model and search usage are metered separately from original receipts. Authenticated actual Work client and search roundtrip verification remains pending.
1.20.50.0
Connector authentication text and contract consistency
Connector 1.20.50.0 fixes garbled Chinese authentication errors and aligns the credential helper’s embedded model contracts with the existing connector contracts. Current default models, permissions, plugins and verified capability scopes are preserved. Update the connector and restart Codex.
Connector 1.20.11.0
Save context settings with any default connection
Choose an HCD model and save its context window and auto-compaction threshold while another connection is the default. Saving keeps your current default connection and chat model. Your settings remain available if you switch connections and apply when you later enable HCD; restart Codex after enabling it.
1.20.10.0
GPT-5.6 Luna: available in the verified safe scope
Text and image input, six reasoning levels, structured output, functions, streaming, cache, background resources and real two-turn Codex use were verified. Medium is the default. A two-response WebSocket chain was tested; mid-response steering is unavailable for this model. Expensive limits remain untested live and budget guards stay active.
1.20.9.0
GPT-5.6 Terra: available in the verified safe scope
Text and image input, six reasoning levels, structured output, functions, streaming, cache, background resources and real two-turn Codex use were verified. Medium is the default. A two-response WebSocket chain was tested; mid-response steering remains closed pending verification. Expensive limits remain untested live and budget guards stay active.
1.20.8.0
GPT-5.6 Sol: available in the verified safe scope
Text and image input, six reasoning choices, structured output, functions, streaming, cache, background jobs and real two-turn Codex use were verified. Medium is the default reasoning effort. WebSocket continuation uses response.create; mid-turn steering is unsupported. Expensive maxima remain untested and budget guards stay active.
1.20.5.0
GPT-6 Sol: available within the verified safe scope
GPT-6 Sol is live within its verified, safely available deployment scope. Text, six reasoning choices, structured output, functions, streaming, persisted turns, cache and background jobs were verified. Authorization, reservation, budget and unsupported-feature guards remain. Maximum context/output, long-context boundaries, large media and high-output compaction were not live-tested under the high-cost exemption; this is not a maximum-size live-test claim.
1.20.7.0
GPT-6 Luna: available within the verified safe scope
GPT-6 Luna is live within its verified, safely available deployment scope. Text, six reasoning choices, structured output, functions, streaming, persisted turns, cache and background jobs were verified. Authorization, reservation, budget and unsupported-feature guards remain. Maximum context/output, long-context boundaries, large media and high-output compaction were not live-tested under the high-cost exemption; this is not a maximum-size live-test claim.
GPT-6 Astra completed acceptance for text, five reasoning levels, structured output, tools, streaming, history, cache, compaction and controlled background resources within the current verified safe scope. Optional hosted features remain restricted, and a JPEG semantic fixture failed. Maximum context/output and other expensive boundaries were not live-tested under the high-cost exemption.
1.20.4.0
DeepSeek V4.1 Flash: native APIs and image files
Text and image understanding, reasoning levels, tools and structured output have been verified, together with native APIs and image-file management. Streaming, conversation continuity and actual usage have been checked. Time-band and public-holiday tariffs apply. Composite performance and value scores remain pending; estimates from older models are not substituted.
1.20.3.0
Text, image and video embeddings with multimodal-embedding-v1
Fixed 1024-dimensional independent vectors support up to 20 mixed inputs, including at most one image and one video. Local images, Data URIs and HCD image cache IDs are supported. Text is limited to 512 tokens per item; the provider truncates longer text, so split it for complete coverage. Original responses, usage and billing receipts are retained. Cancellation may still incur provider charges; requests are not automatically retried. Performance scores remain estimates, not direct benchmarks of this model.
1.20.2.0
Verified multimodal reranking with qwen3-vl-rerank
Connector 1.20.2.0 and Knowledge Studio 1.5.5 support text, images and video, with verified limits of 100 text, 40 image and 4 video documents. Actual usage and original receipts are retained. HCD supports 0 < fps <= 1: the exact service rejects zero despite earlier official [0,1] material, so zero is rejected before a provider call. Conflicting 4000/8000-token boundaries remain upstream-checked and await supplier clarification. Video 493154 points per million tokens remains an estimate for the directory and acceptance reservations only; official supplier pricing is unpublished and actual supplier costs remain pending.
1.20.1.0
GTE text reranking verified for the configured workspace
Connector 1.20.1 and Knowledge Studio 1.5.3 add return_documents, all-result top_n behavior, 500 candidates, original scores, request IDs and actual token usage for gte-rerank-v2. The current service also limits each document to 300000 characters. The supplier marks general service as discontinued; the configured workspace passed live calls, without a long-term availability promise.
1.19.0.0
Qwen3.7 text reranking verified
Connector 1.19.0 and Knowledge Studio 1.5.1 add custom instructions, top_n and full-result retrieval for qwen3.7-text-rerank. Rank up to 500 text documents, each up to 50000 characters. Original scores and actual token usage are retained; errors include traceable request and field details. Update the connector and studio. The price label now identifies input tokens; the model rate is unchanged.
1.18.2.0
Mistral Medium 3.5 joins Text & Code Studio
Connector 1.18.2 adds Mistral Medium 3.5 for text and selected-image questions, JSON output and multi-turn messages. Original stream events return after completion. This model is a studio tool, not a main or subagent model. Codestral and saved defaults remain available. Update the connector and studio, then start a new Codex conversation. Model prices are unchanged; failures retain the HCD request ID.
1.17.1.0
Text & Code Studio joins the connector
Connector 1.17.1 adds an eighth studio with Codestral-2501 for text/code questions and code completion that preserves the surrounding code. Save its default model independently or use one-click studio installation. Use your current HCD Key and model permissions. Stream events return after completion; generated code is not executed or written to files. Failures retain the HCD request ID. Update, install Text & Code Studio, then start a new Codex conversation. Model unit prices are unchanged.
Gateway 1.63.0 · Connector 1.16.0.0
MAI-Thinking-1 is available in Codex
MAI-Thinking-1 now has a Codex bridge. Real Codex checks covered two rounds of function-tool calls and saved encrypted-history continuity. It is text-only, with a 256,000-token context and a 64,000-token maximum output. Reasoning is fixed and has no manual effort control. Function tools, client namespaces and encrypted history are supported. Replies are fully checked and then buffered before delivery; this is not native real-time incremental streaming. Update to Connector 1.16.0.0, refresh the model list and start a new Codex conversation. The native API remains available and the model price is unchanged. New Text and Code Studios are not available in this release.
1.15.2.0
Qwen Flash Character is available in Codex
Connector 1.15.2 adds Qwen-Flash-Character-2026-02-26 after real tool-call and saved-history checks. It retains an 8K context and a 4K output limit. Update the connector and refresh the model list.
Gateway 1.62.0 · Connector 1.15.1.0
Long-running calls can continue while making valid progress
Conversation, reasoning and compaction calls now have separate limits for connection, first valid response, idle time after valid progress and total duration. Ordinary calls use 30/300/180/1800 seconds; complex reasoning uses 30/600/300/3600; compaction uses 30/600/180/1800. Empty heartbeats do not extend a limit, while continued valid progress can pass the former fixed 95-second limit. Cancellation, cleanup and actual usage records remain intact; errors identify the phase, and calls with an unknown result are not replayed automatically. Connector-managed HCD connections now keep stream idle at least 63 minutes for buffered responses, while server-side phase limits can end a request earlier. Save the connection, then start a new task. Higher custom values are preserved. Studio timeouts are unchanged.
Catalogue updated · candidates not integrated
Public points catalogue updated
82 Google version entries join the original public points catalogue, bringing the total to 373. Prices display points only, with candidate status shown in each row. Prices are verified for 62 new entries; 23 have labelled performance estimates and 19 have value estimates within the candidate cohort. Missing data remains pending. Existing search, filters, comparison and export remain available. Actual access follows your account catalogue.
Catalog updated · model review in progress
Google model catalog adjustment and integration progress
The legacy Gemini Developer API catalog entries have been removed; related historical bills remain available. New Google models are being reviewed individually and will be listed only after their actual capabilities and prices are confirmed. For current availability, use the model catalog in your account. Connector 1.15 does not need to be reinstalled.
1.15.0.0
Connector 1.15: one-click setup
Save your API Key, then use One-click setup in the header to apply value mode, refresh chat models, set the default connection, save context settings, install all seven studios, add roles in the current category and save user settings. Existing roles are preserved. Each step reports its result; resolve incomplete steps and run again. Current context and preference selections are used. Fully quit and reopen Codex afterward. Studio calls still require the appropriate Key permissions.
API
Three new native API models
MAI Thinking 1 adds reasoning, function tools and opaque history continuity. Mistral Medium 3.5 supports images, JSON and streamed multiple choices. Codestral 2501 offers Chat and prefix/suffix FIM completion. Use a native-mode API Key with /v1/chat/completions or /v1/fim/completions. These models are not listed in Codex or subagent menus. Connector 1.14 remains current; no installer download is needed.
1.14.0.0
Connector 1.14: document structure and multilingual reranking
Adds Mistral OCR 4, Cohere Parse v5 and Cohere Rerank v4.0 Pro with verified studio delivery and actual usage billing. Document outputs retain tables, source image regions, pages, layout, confidence and structured annotations. Reranking supports text and YAML, result counts, document token limits and priority. OCR accepts PDFs; Parse accepts images only. Annotation and search-unit prices are listed separately. Failed requests are not replayed automatically. Update to 1.14.0.0 and refresh the studios.
1.13.0.0
Connector 1.13: live model choices and 26 role templates
Subagent model choices now follow the signed-in account and active provider, with refresh and visibility controls. Choose among 26 roles in five categories, each with recommended models, alternatives, reasoning effort and an explanation. Missing models and connection changes have repair guidance and an option to inherit the parent model. Native subagents inherit the parent provider and permissions; unsupported controls from 1.12 can be repaired with a backup. Saved changes do not replay running tasks. Update to 1.13.0.0; normal model usage charges apply.
1.12.0.0
Connector 1.12: subagents and custom roles
Version 1.12 introduced subagent settings and four example roles with copying and backups. Some independent permission, plugin and MCP controls shown in that version are not applied by the current native Codex runtime. Upgrade to 1.13 for corrected controls; parent-task permissions apply.
1.11.2.0
Connector 1.11.2: memory preferences, configuration and diagnostics
Memory use, generation and tool-assisted preferences are saved independently. Legacy profiles no longer block saving defaults. Network checks identify failure categories and show a request ID when available. Personal and project settings are preserved during connection updates and restoration.
1.11.1.0
Connector 1.11.1: clearer model counts and cache labels
The catalog contains 284 models, with 141 in the Codex chat picker. Pricing identifies cache hits; usage lists cache writes before cache hits. Prices and actual charges are unchanged.
1.6.0.0 / 1.0.0
Realtime Voice Studio: talk, interrupt and continue in Codex
GPT Realtime 2.1 and Mini support live voice, microphone pause, playback and interruption. Update Connector to 1.6.0.0 and enable the studio and Key permission. In Codex, ask to open HCD Realtime Voice. Click Start to use the microphone; read the transcript afterward. Sessions last up to 8 minutes and charge actual usage, including confirmed interrupted output.
1.5.0.0 / 1.7.0
Image Studio 1.7: three new image models
Adds GPT Image 2.5 Flare, GPT Image 2.5 Sunburst and FLUX.2 flex for generation and editing. Options and usage billing follow the selected model. Update Connector to 1.5.0.0 and Image Studio to 1.7.0, then start a new conversation.
1.4.2.0
Connector 1.4.2: dark appearance by default
The connector now opens in dark appearance when no theme has been saved. Previously selected light, dark or system appearance is preserved. You can change it in the appearance menu at the top right.
1.4.1.0
Connector 1.4.1: enhanced page motion
Flowing light paths, rotating rings, background light bands and section transitions make motion more visible. Text and numbers stay clear. Animations pause in the background, respect system settings, and can be switched off under Connector appearance → Page motion.
1.4.0.0
Connector 1.4: a clearer workspace
Graphite and light themes, sidebar navigation and subtle motion make the connector easier to use. Full grouped numbers improve balance and usage readability. Layout adapts to window size; motion follows system settings. Update from the connector or Download Center. Saved keys, model settings and installed studios are preserved.
1.3.0.0
Direct Zhipu and Kimi models
The API adds 12 models; Connector 1.3.0.0 adds 11 selectable models, including GLM-5.3, GLM-5.3-Flash, Kimi K3 and Kimi K2.7 Code. Context limits, thinking settings and image input follow each model. Kimi K3 supports metered web search. Prices refresh daily; account and Key points remain separate. Update the connector and start a new conversation.
2026-09-12
API keys: compatibility or native mode
Existing and new keys default to compatibility mode. Select native mode on the website for custom code using the model platform’s formats. Authentication, quotas and actual usage billing still apply. Native synchronous and streaming chat interfaces are available according to each model; hosted background jobs and resource management are outside this release. The Connector needs no change.
1.2.1
Connector 1.2.1: DeepSeek and device locations
Adds DeepSeek V4.1 Flash with native image understanding and reviewed reasoning levels. Existing Flash names remain compatible and V4 Pro continues. API key devices now show connector contact and actual request IPs separately, with approximate location, network provider and access time. No extra sign-in; location lookup never blocks service.
1.1.0
Download center and Connector 1.1.0
The new download center brings together Connector downloads, technical documents, product brochures and other software. Public downloads need no sign-in. Connector 1.1.0 uses Windows package version 1.1.0.0 and supports upgrades from 1.0 releases. Compatible features advance the minor version; fixes advance the patch version. An API key is still required after installation.
1.0.106.0
Registered and previously used devices
Connector 1.0.106.0 adds a paginated device list to API keys: computer, system, connector version, first registration, latest contact, IP, calls and Key points. Update and open the connector online, then start a new Codex conversation. Model and studio charges follow original receipts; unattributed history is retained. Registration is free and recent contact does not imply online status.
1.0.105.0
Connector 1.0.105.0 is available
Connector 1.0.105.0 adds 80 chat models and 94 image, speech, document, embedding/reranking and video models. Third-party Alibaba models use qwen-. The Codex menu now contains 131 models. Update the connector and relevant studios, then start a new conversation. Limits and prices follow each model; confirmed supplier costs are recorded separately from the Key multiplier.
1.0.102.0
Six Grok models in Connector 1.0.102.0
Adds Grok 4.6, 4.3, 4.20 Reasoning / Non-Reasoning and 4.1 Fast Reasoning / Non-Reasoning. The Codex menu now includes 53 models. Chat, coding, images, function tools and multi-turn responses were checked. Context limits, reasoning controls and hosted search follow each model’s verified capabilities. Update the connector and start a new conversation. Account and Key points remain separate; cache and context-tier prices apply.
2026-09-10
Complete consumption accounting
Account totals now include models, studios and verified adjustments without charging previous usage twice. Confirmed supplier charges are billed even if delivery fails or is cancelled. Built-in web search is billed at the rate shown for the selected model on the pricing page, in addition to model usage. The Key multiplier affects only Key points. Unconfirmed costs remain pending reconciliation. Refresh the website; no connector update is required.
1.0.97.0
31 Ark models and 3D Studio
Added 15 chat/vision, 5 image, 6 video, 2 embedding and 3 3D models. Third-party Ark models use the doubao- prefix. Connector 1.0.97.0 lists 47 coding models with full names, reviewed reasoning levels and model-specific context limits. Update the desired studios and start a new conversation. 3D Studio requires its own API key permission. Account points reflect actual usage; Key points apply the Key multiplier. Google and model-router remain excluded.
1.0.91.0
Separate account and API Key points
Accounts are charged actual usage points. API Key points equal actual points multiplied by that Key’s coefficient and determine its limits and balance. The website supports editing the Key balance and expiry. Connector 1.0.91.0 shows only the current Key’s points, balance and available balance. Verified legacy overcharges are returned through separate refund entries; Key usage stays unchanged.
1.0.86.0
Document, Knowledge and Video Studios are available
Adds 32 non-Google models: 27 text models, Mistral Document AI, three embedding models and Sora 2. Update to Connector 1.0.86.0 and enable each plugin and its API key permission. Text API: 34 models; Codex menu: 32 models with tools. Phi and Llama support API questions only. Documents: 30 MB and 30 pages. Knowledge: selected local text files, citations and reusable local indexes. Video: 4/8/12 seconds, 720p, matching reference images and completed-video remix. Download videos promptly. Completed usage is billed; cached reads are free. Google and model-router remain excluded.
Available · audio tools
Audio Studio: transcription and speech generation
Install Audio Studio with Connector 1.0.84.0 and enable audio permission for your API key. Four transcription models and GPT-4o Mini TTS are available. Recordings: up to 24 MB and 30 minutes. Diarization also saves speaker segments, TXT, SRT and VTT. Speech: up to 4096 characters, voice, speed and style options, with MP3, WAV, FLAC, Opus, AAC or PCM output. Files stay on your computer; speech is AI-generated. Actual successful usage is billed; local replay is free. Google models and real-time audio chat are excluded.
Available · generation and editing
Image generation and editing are available
Update to Connector 1.0.83.0, update Image Studio in the capability library, then start a new Codex conversation. Use up to 16 local reference images and an optional matching PNG alpha mask. Image 2 supports custom sizes up to 3840 pixels on the longest side; Image 1.5 supports input fidelity, Image 2 uses high fidelity automatically, and Mini has no fidelity parameter. Defaults remain one low-quality PNG. Actual usage determines points; failed requests are not automatically resubmitted.
Restored · fix deployed
Image service restored; release recovery corrected
On 6 September, image generation was unavailable from 00:31 to 09:59 (UTC+8). Our platform release stopped the dependent image service, and the release process failed to restart it. We apologise for the interruption. Service was restored that morning; on 7 September we corrected the recovery process and added checks covering all services. No connector reinstallation is needed. Maintenance may still cause brief interruptions.
Enabled · automatic
HTTPS connection reuse is live
HTTPS connection reuse is enabled for all seven connected chat models, including streaming, and the three image tool models. Repeated requests can avoid reconnecting while a valid connection remains available. First requests, expired connections and long idle periods require a new connection. This reduces connection overhead; it does not change capabilities, prices or usage billing, and no fixed reduction in total generation time is promised. Canceled or failed requests discard their connection without automatically resubmitting generation. No settings change or reinstallation is needed.
Available
GPT Image 1 Mini and GPT Image 1.5 are available
Both models became available on 6 September. Their initial generation-only scope was expanded on 7 September; see the latest editing announcement.
Adds 32 non-Google models: 27 text models, Mistral Document AI, three embedding models and Sora 2. Update to Connector 1.0.86.0 and enable each plugin and its API key permission. Text API: 34 models; Codex menu: 32 models with tools. Phi and Llama support API questions only. Documents: 30 MB and 30 pages. Knowledge: selected local text files, citations and reusable local indexes. Video: 4/8/12 seconds, 720p, matching reference images and completed-video remix. Download videos promptly. Completed usage is billed; cached reads are free. Google and model-router remain excluded. Equipment identities, protocols, operating data, agents and control remain in customer-authorized local, edge or private environments. The public website introduces the service; it is not a cloud equipment-control platform.
