Kling 4, Gemini 4 Argon and Eleven v4 · What Is Actually Live

Quick Summary

• Kling 4.0 Flash is in limited early access for Ultra Yearly subscribers. It is the faster subset, capped at 720p, 8-bit SDR and up to 20 seconds. The full Kling 4.0 model is scheduled for a wider October rollout.
• In Theoretically Media's creator tests, Flash handled dialogue, location continuity and lip sync better than expected, but some performances remained flat and the model introduced unrequested wardrobe and action details.
• The full Kling 4.0 specification adds native 3 to 30-second generation, up to ten keyframes, up to 15 combined reference assets, 1080p or 4K output and 10-bit HDR at the higher resolutions. Those are not all Flash capabilities.
• HeyGen Video is available through its API and launch partners. The advertised $0.01 per second is a temporary 50% launch price for 480p text-to-video or image-to-video through the end of October.
• Gemini 4 Argon has been announced, but it is not broadly released. Google is starting with trusted cyber defenders, then plans access for paid API customers and Google AI Ultra subscribers without publishing a general release date.
• Google lists October 2, 2026 as the shutdown date for the Gemini API endpoint gemini-2.5-flash-image and recommends migration to a newer Gemini image model.
• Ideogram 4.5 is available for precise, repeatable editing with reduced drift. Eleven v4 and v4 Turbo are also available now, with natural-language audio tags for emotion, delivery and non-speech events.

Credit and evidence status

This article is based on Theoretically Media's embedded roundup, published on October 1, 2026, the complete supplied transcript and the official product pages available on October 2. Tim's Kling, Ideogram and ElevenLabs outputs are creator tests. Availability, technical specifications, pricing and retirement dates are checked against first-party sources.

The status labels are essential. Kling 4.0 Flash, HeyGen Video API, Ideogram 4.5, Eleven v4 and Eleven v4 Turbo are available in the forms described below. Full Kling 4.0 is scheduled for October. Gemini 4 Argon is announced and restricted to an initial cohort. Speculation about future Nano Banana, Genie, Veo or assistant releases remains speculation.

Affiliate links are identified in the source list. The article does not use an affiliate relationship as evidence for product quality.

The release-status matrix

Kling 4.0 Flash: limited early access. Ultra Yearly subscribers were the first eligible group. Flash supports text-to-video, image-to-video and Omni Reference at up to 720p, 8-bit SDR and 20 seconds.

Kling 4.0 full model: announced for October. Official specifications list native 3 to 30-second generation, 720p, 1080p and 4K output, 10-bit HDR at 1080p and 4K, up to ten keyframe images and richer multimodal reference control.

HeyGen Video: API available now, with access through OpenRouter, Runware and ComfyUI. HeyGen says its web platform is coming soon. Promotional pricing lasts through the end of October.

Gemini 4 Argon: announced, with initial access through Google's Fairwind Program for trusted cyber defenders. Paid API and Google AI Ultra are named as the next access groups, but Google gives no exact public release date.

Ideogram 4.5: available through Ideogram's creative product and API. Eleven v4 and v4 Turbo: available through ElevenCreative, ElevenAgents and the ElevenLabs API. Gemini 2.5 Flash Image: listed for API shutdown on October 2, 2026.

Kling 4 is here, but Flash is not the full model

Kling's launch materials make a distinction that is easy to lose in a headline. Kling 4.0 Flash is the early-access model available first. It offers the core text-to-video, image-to-video and Omni Reference experience, but its output is limited to 720p, 8-bit SDR and up to 20 seconds.

The full Kling 4.0 model is the larger October release. Its official specification adds native generation from 3 to 30 seconds, 1080p and 4K options, 10-bit HDR at the higher resolutions, first and last frames, multiple keyframes and a larger reference system. Video extension up to two minutes is described as coming soon rather than available at launch.

That difference changes how the early tests should be read. A soft 720p result does not prove that the full model lacks detail. A missing multi-keyframe control in Flash does not contradict the full-model roadmap. The correct early question is whether Flash improves the everyday generation loop enough to justify testing before the wider rollout.

Five Kling tests and what they reveal

Text-to-video, jaywalking ticket: the scene holds together, but the police officer performs an unrequested gesture. This is a prompt-adherence warning. Plausible action is not the same as requested action.

Image-to-video, FBI diner: the model has to invent the waitress when she turns from the source image. Tim reports that the identity and visual style remain coherent while the model completes the dialogue with time to spare. Treat that as one successful creator example, not a measured consistency rate.

Omni, Flamethrower Girl birthday: character identity and room geography remain recognizable, including the sink position, but the scene ignores secondary physical consequences such as a melting candle. This shows why spatial continuity and causal physics need separate review criteria.

Omni noir, Tom Malloy's office: eyelines and dialogue work, but Kling changes the detective's jacket to gray and the acting feels flat. The result is visually usable coverage with a costume-continuity failure and a performance problem.

The most encouraging creator observation is lip sync. Tim sees the drift that affected his Kling 3.0 dialogue tests as substantially improved in these Flash generations. This should be retested across longer speech, accents, faster delivery and overlapping dialogue before calling it solved universally.

Why keyframes plus Omni may be the stronger workflow

Tim's vampire-hunter test combines planned frames with Omni references. His working theory is that keyframes give the model clearer performance and composition targets, while Omni references preserve identity and other recurring details.

Kling's official guidance describes the same division of labor. Keyframes define when a visual state or story beat should change. Omni Reference defines what should remain consistent, including characters, products, movement, camera language, style and voice.

The full Kling 4.0 model supports up to ten keyframe images and up to 15 combined reference assets. The combined reference pool can include up to ten images, five videos with no more than 30 seconds total, subjects and an authorized voice reference. Individual limits still apply within the combined total.

Use fewer references when possible. Give each one a single job, such as character identity, product design, camera movement or final composition. Extra inputs that disagree can reduce control instead of improving it.

HeyGen Video costs one cent per second, with conditions

HeyGen Video is HeyGen's first general-purpose video model, built from MiniMax H3 and post-trained by HeyGen. Tim had not tested it when recording, so the video provides a release and pricing summary rather than a quality verdict.

The $0.01 per-second headline applies to 480p text-to-video and image-to-video during a 50% launch promotion through the end of October. The corresponding standard rate from November is $0.02. Promotional 768p costs $0.015 per second and rises to $0.03.

Reference-to-video costs more because billing covers both the input and output seconds. The promotional rates are $0.02 per second at 480p and $0.03 at 768p, rising to $0.04 and $0.06 respectively after the promotion. Image and audio references do not add to that billing calculation.

The API is live, with launch access through OpenRouter, Runware and ComfyUI. HeyGen says web-platform access is coming soon. Calculate the total for resolution, mode, source-video duration, output duration, retries and approved clips before comparing it with another provider.

Gemini 4 Argon is announced, not generally released

Google announced Gemini 4 Argon on September 30. The initial rollout is to trusted cyber defenders through the Fairwind Program while Google tests safeguards and participates in pre-release review. The company says broader access will start with paid API customers and Google AI Ultra subscribers, but it does not publish a date.

Google positions Argon for long-horizon coding, professional knowledge work and defensive cybersecurity. It lists a one-million-token context window and an introductory API price of $2 per million input tokens and $10 per million output tokens. Google says the later standard price will be $4 and $20.

This is not a Gemini app launch that every creator can test today. Until normal access opens, the most useful facts are the intended workload, rollout order, price plan and published evaluations. Claims about how Argon changes creative workflows remain projections.

Read the Argon benchmark charts with caution

Google reports a 91.7 score on LVBench for long-video understanding and presents Argon as state of the art on that evaluation. That result is directly relevant to creators who want a model to inspect long edits, find moments, compare scenes or reason across a full film.

Theoretically Media counts Argon ahead of GPT-6 Astra and Claude Opus 5.5 on 12 of the 18 benchmark rows shown in Google's launch material. This is a reading of vendor-selected launch comparisons, not an independent universal ranking.

Vals independently lists Argon first on its overall index at the time checked, while also showing that other models lead individual tasks. That is the more useful interpretation: a strong aggregate result can coexist with second, third or lower placement on a workflow that matters to you.

Before changing a production stack, test the exact job: long-video retrieval, storyboard criticism, code migration, legal drafting or image analysis. Record prompt, context size, tool access, latency, cost and reviewer score. A benchmark without the operating conditions is not a buying decision.

What comes after Argon is still speculation

Tim asks whether a stronger flagship model will power a future Nano Banana 2 Pro, a Genie update, a renewed Veo push or a more capable agentic assistant. These are reasonable product questions, but the cited Argon announcement does not confirm those releases.

Keep them in a watchlist rather than a production roadmap. Do not promise a client features, dates or integrations that Google has not announced. The verified statement is narrower: Argon is Google's new frontier model, restricted access comes first and the next named audiences are paid API and Google AI Ultra customers.

Gemini 2.5 Flash Image reaches its API shutdown date

Google's deprecations table lists October 2, 2026 as the shutdown date for gemini-2.5-flash-image, the API model widely known as the original Nano Banana. Google recommends moving to a newer Gemini image endpoint.

If the model is still present in a consumer interface or a third-party platform, that does not mean the old Gemini API endpoint remains safe for production. Product surfaces and API model IDs have different lifecycles.

Migration checklist: inventory every hard-coded model ID, save representative inputs and accepted outputs, run the same cases on the replacement, compare typography, identity, edit preservation, latency and cost, then update fallbacks and monitoring. Do not wait for the first failed production request.

Ideogram 4.5 focuses on precise editing

Ideogram calls 4.5 its most precise edit model. Its core promise is reduced drift across repeated edits: change one requested area while preserving shape, texture, color and untouched pixels more reliably than a full regeneration.

Tim tests it on an older image containing garbled text, corrects the lettering and replaces a profile image with Flamethrower Girl. His initial result is positive, but he also says he has not spent enough time with the model for a final verdict.

The official API adds a Precise Edit route, optional masks, up to four reference images and high-resolution editing without automatically downsizing the original. Those controls make 4.5 especially relevant for packaging, signage, localization, product colorways and iterative campaign revisions.

Ideogram 4.0 is available as an open-weight model. The 4.5 page checked for this article presents the newer editing model through Ideogram's product and API but does not provide a 4.5 weights download. Treat the video's open-weights comment as a reported future direction until Ideogram publishes the release terms.

Eleven v4 brings direction into the script

ElevenLabs released Eleven v4 for high-quality expressive speech and Eleven v4 Turbo for low-latency use. Both are available through ElevenCreative, ElevenAgents and ElevenAPI, and ElevenLabs offers free-account access for initial generation.

The main workflow change is the audio tag. Short natural-language instructions in square brackets can guide emotion, delivery and audible events. Official examples include [curious], [whispers], [laughs], [sighs], [leaves rustling] and [applause]. Punctuation and surrounding context still affect the result.

Tim's high-fantasy watchtower test combines narration, two characters, cold-weather performance and an ominous fire effect. The result demonstrates how one script can contain performance direction instead of requiring every line to be generated and assembled separately.

ElevenLabs warns that tags are not perfectly reliable and can be confused with sound-effect requests. Start with the fewest tags needed, test the exact voice, keep dialogue unchanged and regenerate only the line that fails. Use cloned voices only with clear authorization from the speaker.

A practical test plan for this release wave

Kling: run one dialogue scene in Flash today, then repeat the same input on the full model when access opens. Score identity, wardrobe, eyelines, lip sync, physical causality, audio, resolution and cost per accepted second.

HeyGen: compare 480p and 768p on the same short prompt. Include launch and standard pricing in the spreadsheet. Test text-to-video separately from reference-to-video because the billing and control mechanisms differ.

Ideogram: begin with an approved image and complete five sequential edits. After every pass, compare untouched pixels, typography, product shape, skin, palette and resolution with the original.

ElevenLabs: use one short dialogue with restrained emotional tags. Compare v4 and v4 Turbo for expressiveness, identity, latency and tag obedience. Keep the untagged version as a control.

Gemini: migrate 2.5 Flash Image now. For Argon, prepare a representative evaluation set but wait for actual access before publishing conclusions about creative performance.