A model wave is a re-pricing event.

Between 16 and 24 July 2026, five models shipped and one access boundary moved. Moonshot AI released Kimi K3 in mid-month, Alibaba previewed Qwen3.8-Max on 19 July and announced Qwen-Image-3.0 around the same days, Google shipped Gemini 3.6 Flash on 21 July, and Anthropic launched Claude Opus 5 on 24 July. The sixth event is different in kind and belongs in the list anyway: on 20 July, Anthropic changed which Claude subscriptions include Claude Fable 5. Not a model. It moves the same routing variable the models move, which is the point of this piece.

The reflex is to ask which model is best. That question has no operational answer, because nobody runs a model. They run a set of routes. A team with agents in production already decided that some work goes to a fast cheap lane, some to a mid lane, some to the strongest thing available. Every one of those choices was made under specific prices, context limits, output caps, latency profiles, retention terms and access rules. A wave moves several of those inputs at once, in different directions, for different vendors. What it invalidates is not your model choice. It is the arithmetic under your model choice.

So the question that produces work is narrower: for each class of work you already run, what did this week change about the cost of that lane, the risk of that lane, and the cost of leaving it. Sometimes the answer is nothing, and nothing is a legitimate finding worth a dated line in a table.

We publish the method with the conclusion. Every figure below comes from a vendor's own documentation, a vendor's own announcement, or an independent measurement page, each fetched directly on 25 July 2026. Where a page could not be fetched or a number was not published, we say so and we do not estimate.

The most useful finding is a negative one. Of the five models, exactly one published numeric benchmark results we could extract as text: Gemini 3.6 Flash, on its DeepMind model card. Anthropic's Opus 5 announcement presents every result as a chart, and we could not pull a single numeric score out of it. Moonshot's Kimi K3 post carries a section headed as a full benchmark table whose numbers are rendered client-side and appeared in none of our fetches. Alibaba published no benchmark at all, for either Qwen release. Four of the five models in this wave shipped no benchmark number we could extract from a fetched page. Any score we printed for those four would be a chart transcription or an aggregator's copy, and we print neither.

Second finding: the dates conflict. Kimi K3 is dated 17 July 2026 by Moonshot's own blog and by Forbes, and 16 July 2026 by OpenRouter and by Simon Willison. The likeliest explanation is an announcement in UTC+8 straddling a date boundary. We do not adjudicate; we write mid-July. Qwen-Image-3.0 is messier: press converges on 21 July, the official blog page metadata carries 16 July, and Alibaba Cloud's English API reference reports a last update of 20 July. No primary source states a release date in body text at all.

Third finding: two of the five were announced with open-weights intentions, and neither had public weights when we checked. Kimi K3's licence is stated on no page we could fetch, and with no public weights there is no public licence file to read either. Anthropic's Opus 5 system card exceeded our fetch size limit, so nothing in it is verified here, including the safety level that secondary outlets report. Everything in this piece is stamped 25 July 2026, and some of it will be stale within days.

The extraction lane covers classification, extraction, normalization, formatting and schema conformance. Its failure mode is silent and well-formed: a wrong field, a dropped row, a plausible category on an implausible input. The output looks correct, so reading it is the wrong check. Assertions, fixtures and sampling are the right one, and that is why this lane tolerates a cheaper model than intuition suggests. The safety net is not the model. It is the check.

The implementation lane covers routine code, tests, migrations of a kind you have done before, documentation and small-diff review. Its failure mode is loud: it breaks CI or gets caught in review. We price this lane in review minutes rather than tokens, which means a model that produces slightly worse code and much shorter diffs can win outright.

The frontier lane covers architecture, ambiguous debugging, unfamiliar migrations, security-sensitive synthesis and any decision where being wrong costs more than every token you will ever spend on it. Its failure mode is the dangerous one: a confident, coherent, internally consistent answer that is wrong in a way the reviewer lacks the context to detect. This lane is priced by consequence. Token cost is not the quantity that decides it.

The multimodal lane is really two lanes. Understanding takes images, audio, video or documents in and returns text. Generation returns an artifact. They have different vendors, different failure modes, different cost structures and, this month, very different licensing clarity. Treating them as one lane is a routing mistake, and this wave makes the split easy to see.

Google announced Gemini 3.6 Flash on 21 July 2026; its docs state it is generally available and ready for production, under the API model ID gemini-3.6-flash. Standard pricing is $1.50 per million input tokens and $7.50 per million output tokens, with output including thinking tokens. Batch is $0.75 and $3.75, flex identical to batch, priority $2.70 and $13.50. Context caching costs $0.15 per million input tokens plus $1.00 per million tokens per hour of storage, so the storage term, not the read rate, is the number to model. The context window is 1M input with a 64K output limit. The knowledge cutoff is March 2026, with the model card noting some domains limited to January 2025. Inputs may be text, images, audio or video, with PDFs listed on the DeepMind Flash page. Output is text only.

The four price tiers are the actual product: work that can wait should not pay the synchronous rate, and batch is half of standard. Google also claims 3.6 Flash reduces output token usage by 17% against 3.5 Flash, per the Artificial Analysis Index. Read that carefully, because it is the number most likely to be misquoted this month: it is a reduction in tokens consumed, not a price cut. For output-billed work it is arguably the more interesting number, since consumption is the term you cannot negotiate. Secondary reports also describe a cut to the output price, but we could not verify the prior figure on any Google primary source, so we neither print it nor lean on it. Google further cites Datacurve reporting up to 65% output token reduction on the DeepSWE benchmark, and partner figures from Harvey (12% faster task completion) and JetBrains (10 to 20% coding performance gains).

The 64K output cap is the specification that decides whether a job belongs in this lane. For extraction it almost never binds. For any job whose artifact exceeds 64K tokens it is a hard wall, and no amount of intelligence compensates. For contrast: Claude Opus 5 emits up to 128k tokens synchronously and up to 300k on the Message Batches API behind a beta header, and Kimi K3 defaults to 131,072 output tokens and can be configured up to 1,048,576. Three 1M-context models whose output ceilings differ by a factor of sixteen. Context window is the number everyone quotes. Output cap is the number that routes work.

Two independent measurements from Artificial Analysis point in opposite directions and both matter. It measured 237.1 output tokens per second, ranked first of 190 models, and a time to first token of 14.73 seconds, which it describes as somewhat higher than average. A person watching a cursor experiences the first token; a queue draining output experiences the rate. The same model ranks first on one measurement and below average on the other. Artificial Analysis scores it 50 on its Intelligence Index v4.1, ranked 26th of 190 against a median of 32.

Anthropic announced Claude Opus 5 on 24 July 2026 at $5 per million input tokens and $25 per million output tokens. That is exactly what Claude Opus 4.8 cost, and exactly half of Claude Fable 5 at $10 and $50. Anthropic's framing is that Opus 5 comes close to the frontier intelligence of Fable 5 at half the price, and that it delivers greatly improved performance for the same cost as its predecessor. Both statements describe the same price list. If your frontier lane runs on Fable 5, the vendor is now arguing that the lane costs half. Whether it does on your workload is what a canary is for, because a spec sheet does not carry pass rates, retry burden, or review minutes.

There are no numeric benchmark scores to check. The announcement presents everything as charts, and the named evaluations are Frontier-Bench v0.1, CursorBench 3.2, the AA Coding Agent Index, ARC-AGI 3, GDPval-AA v2, OSWorld 2.0, HLE, Zapier AutomationBench and DeepSearchQA. What we could verify is Anthropic's own wording about relative position, with no absolute numbers behind it: on Frontier-Bench v0.1 it surpasses all other models and more than doubles Opus 4.8's performance at a lower cost per task; on CursorBench 3.2 at max effort it is within 0.5% of Fable 5's peak score at half the cost; on ARC-AGI 3 its score is three times the next-best model; on Zapier AutomationBench its pass rate is around 1.5x the next-best model for the same cost; on OSWorld 2.0 it surpasses Fable 5's best result at just over a third of the cost. Four of those five claims are cost-normalized. The vendor is arguing on the same axis this piece argues on, which is interesting, and unfalsifiable from outside without the numbers, which is the problem.

The effort parameter is the real routing dial, and it means the price list is not the price. Opus 5 supports five levels: low, medium, high, xhigh and max, with high as the API default on both the Claude API and Claude Code, set through output_config in the Messages request. Thinking cannot be disabled at xhigh or max; a request that tries returns a 400. Anthropic also notes that on Opus 5, changing effort does not reliably shorten the visible response, and that you should prompt for length instead. Effort moves token consumption, and token consumption is what you are billed for. A cost model that treats $5 and $25 as constants is modeling the list, not the bill.

Two line items on the rest of the surface move the bill by multiples; the others are small constants worth knowing. Caching first: cache hits and refreshes cost $0.50 per MTok against $5.00 fresh, and a five-minute cache write costs $6.25, a premium a single reuse inside the window more than repays. The one-hour write costs $10.00 and needs at least two reads against the same prefix to come out ahead. Effort is the second, covered above. The small constants: batch at $2.50 and $12.50; fast mode at $10 and $50 for around 2.5 times default speed, in research preview, first-party Claude API only, not available on Claude Platform on AWS, partner clouds, or the Batch API; the full 1M window billed at standard rates with no long-context surcharge on Claude 4.6 and later; tool use adding 286 input tokens of system prompt overhead for tool_choice auto or none and 406 for any or tool, with the Bash tool adding a further 325; and Claude Managed Agents at $0.08 per session-hour on top of tokens. None of that is exciting. All of it shows up on the invoice.

From 20 July 2026, Claude Fable 5 is included at no extra cost on Max, on premium Team seats, and on premium seat-based Enterprise seats, capped at 50% of weekly usage limits. That cap is not additive: other models draw from the same weekly pool, and the total is still the weekly limit. Pro and Team Standard seats are not included and pay through usage credits after a one-time credit. Anthropic's help centre says only that a one-time credit exists; the @claudeai account, as quoted by Simon Willison, gives the figure as $100. The prior free-access promotion ended 19 July 2026 at 11:59:59 PM PT. One wording matters here: the help-centre article never uses the word permanent. It states a change effective 20 July with no end date, and the 'permanent' in the headlines is the journalists' word, not a quote. An absent end date is not a commitment, and the difference between those two readings is the whole question of whether you can build on this.

The instability is itself the data. Anthropic revised Fable 5's plan access at least four times between 9 June and 20 July 2026, and its own stated reason is that demand proved hard to predict, so access was extended in stages as capacity was secured. Separately, Fable 5 was forced offline on 12 June 2026 by a US export-control directive and returned on 1 July. Independent reporting describes the practical effect of the July change as 50% of already-reduced limits, following a roughly one-third cut to regular usage when the bonus phase ended. None of that says anything about the model's capability, and all of it is a routing input. A lane whose commercial terms moved four times in six weeks, and which a government directive took offline once, should not be the only lane a team is able to run.

Two constraints in this release beat capability outright. Fable 5 is designated a Covered Model with 30-day data retention and is not available under a zero-data-retention agreement. For a regulated workload that is dispositive, and no benchmark can argue with it. The second is quieter and costs teams real money: Fable 5 uses the tokenizer introduced with Claude Opus 4.7, and the same text produces roughly 30% more tokens than on models released before 4.7. Dollars per million tokens is therefore not a comparable unit across vendors, or even across generations of one vendor. Price your lanes on your own corpus, in tokens you counted yourself, or you are comparing two different rulers.

Kimi K3 is real, hosted, and technically the most interesting thing in the wave. Moonshot's blog describes a 2.8 trillion parameter model using a sparse mixture of experts that activates 16 of 896 experts per token, an attention design named Kimi Delta Attention with attention residuals, and quantization-aware training with MXFP4 weights and MXFP8 activations. vLLM's preview write-up corroborates the expert configuration and the MXFP4 weights, and describes the attention as KDA-dominant linear attention with periodic full-attention layers. Inputs are text, images and video files per the API docs. Context is 1,048,576 tokens. Pricing is $3.00 per million input tokens on a cache miss, $0.30 on a cache hit, and $15.00 per million output, flat, with no tiering by context length. Reasoning is always on, with effort levels low, high and max, defaulting to max. API access unlocks after a minimum $1 top-up. Moonshot positions it for long-horizon coding and end-to-end knowledge work.

What is missing is the part that matters for resilience. As of 25 July 2026 the moonshotai organisation on Hugging Face lists no K3 repository; the newest entries are K2.7-Code and older. Moonshot's blog commits to releasing full weights by 27 July 2026, and vLLM restated that as still pending on 22 July. Third-party repositories matching the name exist and are not Moonshot's. No page we fetched states a licence, and with no public weights there is no public licence file to read. vLLM's support is preview-stage and requires large-scale expert parallelism. None of this stops a team calling the hosted API today; the model ID, prices and limits are published and the lane is usable. What it stops is the thing open weights are for: running it, pinning it, and no longer caring what the vendor does next. That part is a date on a calendar, two days after this piece publishes.

Qwen3.8-Max-Preview is thinner still. Alibaba's Qoder documentation states 2.4 trillion parameters. It was previewed on 19 July 2026 at the World Artificial Intelligence Conference in Shanghai. Access runs through Alibaba's Token Plan subscription, Qoder and QoderWork, and Token Plan is available only in the China (Beijing) region. There is no published per-token price; billing is credits-based, with the preview discounting the billing coefficient from 0.5x to 0.05x, and to 0.01x between 22:00 and 08:00 UTC+8. Alibaba published no model card, no benchmark table, no activated-parameter count and no licence, and we confirmed the absence directly: the Model Studio specification tables contain no row for this model. The only performance claim Alibaba made is a single post from the Qwen team describing it as second only to Fable 5, with nothing published behind it. Weights are promised soon, with no date and no named licence. One trap worth flagging: a context-window figure for this model circulates widely and appears on no Alibaba page we fetched; the 1M figure in those tables belongs to Qwen3.7-Max, a different model. We state no context window for Qwen3.8-Max-Preview, because we do not have one.

Two vendors announced very large models with open-weights intentions inside one week, and neither has shipped public weights. Provider resilience is not a second API key, and it is certainly not a promised weight drop. It is three concrete things: a tool contract portable across providers, a pinned replay set that defines success independently of any model, and a second lane you have actually run under production load at least once. Build all three as if the second lane will be needed. Then treat every announced weights release as a diary entry to verify, not a capability you already possess.

On the understanding side, this wave was generous. Gemini 3.6 Flash accepts text, images, audio and video, with PDFs listed on the DeepMind Flash page, and returns text only. Kimi K3 accepts text, images and video files, with images supplied as base64 or an ms:// file identifier; its docs state inputs and make no statement about output modality beyond text responses being the default. Qwen3.8-Max-Preview is reported to handle text, images, video and document understanding, but that comes from secondary coverage only and we could not confirm it on any Alibaba-owned page. Understanding routes like the extraction lane with a larger bill: emit a structured object, assert on the object, and never let a reviewer's impression of the prose stand in for a check.

Generation is a different discipline. Qwen-Image-3.0 is exposed on Alibaba Cloud Model Studio as qwen-image-3.0-pro, handling text-to-image and image-to-image in one model, with image-to-image taking one to three reference images plus a single instruction. It is in limited preview and requires an approved application through the Model Gallery. It serves from two regions, China (Beijing) and Singapore, with separate keys and endpoints that cannot be used interchangeably. Verified API constraints: total output pixels between 512x512 and 2048x2048, PNG output, n between 1 and 6 images with a default of 1, prompt_extend defaulting to true, watermark defaulting to false, seed in the range 0 to 2147483647. Input images may be JPG, JPEG, PNG, BMP, TIFF, WEBP or GIF, recommended between 384 and 2048 pixels per side, up to 10MB. Requests are single-turn with exactly one user message.

The vendor claims are the interesting part and none of them is checkable. Qwen's blog states support for prompt input up to 4.5k tokens, precise rendering of text as small as 10px, native rendering of 12 languages, and over 100 artistic styles; it offers an example of a three-by-three grid of infographics generated in a single pass from a 3.7k-token prompt rather than stitched together. There is no benchmark table, no parameter count, no architecture description, no technical report and no licence statement on the blog or in the API reference. The price is listed as free for a limited time in both regional rate tables, which is not a price. For scale, the same tables put the prior generation, qwen-image-2.0-pro, at ¥0.5 per image in Beijing.

Two of those facts are pipeline design rather than trivia. Generated image URLs expire after 24 hours, so any system that retains images must persist them on receipt; a pipeline that stores the URL has a 24-hour fault built into it that will not surface until someone opens an old record. And no licence is published for the model or for commercial use of its outputs. Our policy is blunt on this: we do not route client deliverables through a generation lane whose usage terms we cannot produce in writing, and nothing published for Qwen-Image-3.0 clears that bar yet. The Qwen image line's last open weights were Qwen-Image-2512, Apache 2.0, dated 31 December 2025. Version 3.0 is closed, invite-gated, and unlicensed in public. That is a supply-chain decision before it is an aesthetic one.

Write the policy as a table of work classes, not model names. For each class, record six fields: the default lane, the escalation trigger, the escalation lane, the hard constraints (retention, region, output cap, licence), the fixture set that proves the lane works, and the date the lane was last re-priced. A policy that lists model names without work classes is a preference. A policy that lists work classes without dates is a memory. The dates are what let you answer, in six weeks, whether a decision was made under conditions that still hold.

Then apply the wave. Extraction and formatting default to a cheap lane, and if that lane is Gemini 3.6 Flash, route asynchronous volume to batch at $0.75 and $3.75 rather than standard at $1.50 and $7.50, and check that no job in the class needs more than 64K output. Routine implementation stays wherever your review-minute cost is lowest; nothing verified in this wave forces a move, and a lane that produces shorter diffs beats a lane that scores better. Frontier work routes to the strongest lane you can afford per task, and if that lane was Fable 5, run a canary on Opus 5 at $5 and $25 with effort at xhigh before treating the halved list price as a halved cost. Multimodal understanding joins the extraction lane with harder assertions. Multimodal generation stays out of client work until usage terms exist in writing.

Measure five things. Cost per accepted task, not cost per million tokens, because effort levels and tokenizers make per-token comparisons meaningless across vendors. Escalation rate. Escalation win rate, meaning how often the frontier lane actually changed the answer. Review minutes per accepted task. And the date of your last failover rehearsal. If escalation rate is high and escalation win rate is low, the lane is fine and the trigger is miscalibrated; fix the trigger, not the lane. What this piece cannot hand you is thresholds. Those come out of your telemetry, and any consultant who quotes one without seeing your telemetry is guessing.

The standing rule is that you re-price when an input moves, not when a model launches. Price, context limit, output cap, access terms, retention terms, region: any of those moving opens the table. The rule needs one honest amendment: a capability claim at unchanged list price is not a re-route trigger, but it is a canary trigger, because cost per accepted task can fall while the price list holds still, through fewer retries and shorter reviews. The re-route still waits for the canary's numbers, and the canary's arithmetic must include the cost of switching itself: integration work, prompt portability, cache loss, and reviewers learning a new failure profile. By our reading of this week: one release re-priced the cheap lane, one re-priced the frontier lane pending canary, one moved an access boundary without touching a model, one added a hosted lane whose weights and licence are promises with a date attached, and two we cannot act on at all, one gated to a region and subscription we do not hold and one invite-gated with no licence. The first of those promises comes due on 27 July. Write down what moved, what you did about it, and what you could not verify, and date all three.