Three AI labs in China shipped flagship-class language models within a three-week span this fall, and the numbers behind them tell a messier story than the “China is catching up” headlines suggest. Alibaba’s Qwen3.8 Max Prime, Zhipu AI’s GLM 5.3 Prime, and StepFun’s Step 5 Preview all landed between September 23 and October 8, 2026, each chasing the same agentic-coding market that OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Opus 5.5 currently lead. On paper, all three look similar: 1-million-token context windows, function calling, streaming, JSON mode, reasoning support. In practice, the gap between the best and worst of the trio is as wide as the gap separating any of them from a US frontier model.
This comparison pulls verified pricing, scores, and release data from three independent trackers, lines up all three new entrants against each other and against GPT-6.1 Sol, Claude Opus 5.5, Claude Sonnet 5.5, Mistral Large 4, and Grok 4.7, and works out where each Chinese AI model actually earns its keep. The short version: GLM 5.3 Prime posts the strongest composite score of the three at a mid-tier price, Step 5 Preview is the newest and cheapest, and Qwen3.8 Max Prime is both the most expensive and the lowest-scoring, despite coming from the same family that topped independent charts earlier in 2026.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Three Chinese AI Labs Ship New Flagships Within Three Weeks
The timeline matters here. Zhipu AI’s GLM 5.3 Prime, a high-speed variant of the base GLM-5.3 model, went live on September 23, 2026, the same day Alibaba’s Qwen3.8 Max Prime appeared as a separate, higher-throughput SKU of Qwen3.8 Max. Less than three weeks later, on October 8, 2026, StepFun’s Step 5 Preview arrived, built on a sparse mixture-of-experts architecture with 27 billion active parameters out of 600 billion total, according to model-tracking site LM Market Cap. All three are tracked as part of a broader wave the site has catalogued as 466-plus models released in 2026 alone, with eight new entries in October through the 12th.
What makes this trio worth comparing as a set, rather than as three unrelated releases, is that each one is explicitly positioned for the same workload: long-context, tool-using, agentic coding sessions where a model has to hold a large codebase in memory and call functions repeatedly without losing the thread. Wikipedia’s “Comparison of generative AI models” page, last updated October 10, 2026, places all three in its language-model table alongside GPT-6.1 Sol, Claude Opus 5.5, Gemini 4 Argon, and Grok 4.7, scored on the Artificial Analysis Intelligence Index, Arena.ai’s Text Rating, the Epoch Capability Index, and the Vals Index. That is four independent scoring methodologies looking at the same set of Chinese AI model releases, which is a rare case of overlapping coverage for models this new.
The pattern that emerges across every one of those scoring systems is consistent even where the absolute numbers differ: GLM 5.3 and Qwen3.8 Max cluster together near the top of the open-weight tier, Step 5 Preview sits a notch below them, and all three trail the US frontier tier (Claude Opus 5.5, Claude Sonnet 5.5, GPT-6.1 Sol, GPT-6 Astra) by a visible but no longer enormous margin. That is the real story: not that a single one of them has “caught up,” but that the gap has compressed into single digits on some scales while individual pricing and throughput choices still vary by 4x to 12x across the three new releases.
Meet the Contenders: Qwen3.8 Max Prime, GLM 5.3 Prime, and Step 5 Preview
Qwen3.8 Max Prime (Alibaba)
Qwen3.8 Max Prime is a higher-throughput variant of Qwen3.8 Max, Alibaba’s flagship, served as a separate SKU at a higher price point, according to its listing on LM Market Cap. It accepts text, image, and video input, carries a 1.0-million-token context window with a 131,100-token maximum output, and supports vision, function calling, streaming, JSON mode, and reasoning. Pricing runs $4.00 per million input tokens and $12.00 per million output tokens, among the priciest options in this comparison. On LM Market Cap’s composite score, Qwen3.8 Max Prime lands at 40 out of 100, ranked 243rd out of 437 tracked coding models, putting it in roughly the 56th percentile. The base Qwen3.8 Max (the “0902” build dated September 2, 2026 in Wikipedia’s tracking table) scored 45 on the Artificial Analysis Intelligence Index, with an Arena.ai Text Rating of 1,483 and an Epoch Capability Index of 156.4, but a steep $5.41 cost-per-task figure and a median generation speed of just 36 tokens per second, the slowest of the three new releases by a wide margin. Alibaba lists the model as proprietary.
GLM 5.3 Prime (Zhipu AI)
GLM 5.3 Prime is the high-speed variant of Zhipu AI’s GLM-5.3, inheriting its full capabilities while delivering 1.5x to 2x the output throughput through inference acceleration, per LM Market Cap’s model page. It supports text input and output with a 1-million-token context window and a 131,000-token maximum output, plus function calling, streaming, JSON mode, and reasoning. Pricing sits at $2.80 per million input tokens and $8.80 per million output tokens for the Prime tier, while the base GLM-5.3 model is dramatically cheaper at $0.08 per million input tokens and $4.80 per million output tokens. Both variants carry the same composite score of 79 out of 100 on LM Market Cap, ranking in the top quartile of the site’s 437 tracked coding models. On the Artificial Analysis Intelligence Index, as tabulated by Wikipedia, base GLM-5.3 scores 45, tied with Qwen3.8 Max, with an Arena.ai rating of 1,478, an Epoch Capability Index of 155.6, a $2.01 cost-per-task, and a median speed of 83 tokens per second. Crucially, Wikipedia lists GLM-5.3’s license as MIT, meaning the weights are genuinely open, unlike its two rivals in this comparison.
Step 5 Preview (StepFun)
Step 5 Preview is StepFun’s flagship model for agentic work, built on a sparse mixture-of-experts architecture with 27 billion active parameters and 600 billion total parameters, according to LM Market Cap. It supports vision, function calling, streaming, JSON mode, and reasoning across a 1.0-million-token context window with a 64,000-token maximum output, the smallest output ceiling of the three. Pricing is the lowest of the group at $1.00 per million input tokens and $2.70 per million output tokens. LM Market Cap scores it 75 out of 100, ranked 132nd of 437 coding models (top 30 percent), with the listing dated October 8, 2026. Artificial Analysis evaluated the model slightly earlier, publishing an Intelligence Index score of 44 on September 18, 2026, alongside an Arena.ai Text Rating of 1,447 and a $1.03 cost-per-task figure, the cheapest of the three on that metric, with a median generation speed of 87 tokens per second, the fastest of the trio. Wikipedia lists its license as proprietary and dates its broader release to August 2026, ahead of the specific Preview build most trackers picked up in September and October.
One nuance worth flagging: Alibaba also sells a much smaller Qwen3.8 27B model, a completely different product from Qwen3.8 Max Prime despite the shared “Qwen3.8” branding. Artificial Analysis evaluated that smaller model separately and recorded Intelligence Index scores between 20 and 28 depending on reasoning effort, far below the Max Prime tier. Anyone comparing “Qwen3.8” scores across sources needs to check which SKU is actually being measured, since the family spans a huge capability range under one name.
Full Specs Comparison: 11 Models Side by Side
| Model | Developer | Context Window | Max Output | License | LM Market Cap Score | Release Date |
|---|---|---|---|---|---|---|
| Qwen3.8 Max Prime | Alibaba | 1.0M tokens | 131.1K tokens | Proprietary | 40/100 | Sep 23, 2026 |
| GLM 5.3 Prime | Zhipu AI | 1.0M tokens | 131K tokens | MIT (base model) | 79/100 | Sep 23, 2026 |
| Step 5 Preview | StepFun | 1.0M tokens | 64K tokens | Proprietary | 75/100 | Oct 8, 2026 |
| GLM-5.3 (base) | Zhipu AI | 1.0M tokens | 131K tokens | MIT | 79/100 | Aug 2026 |
| Mistral Large 4 (Preview) | Mistral AI | 1.049M tokens | 262K tokens | Proprietary | 72/100 | Oct 6, 2026 |
| Claude Opus 5.5 | Anthropic | 1.0M tokens (tested) | n/a | Proprietary | 95/100 | Sep 22, 2026 |
| Claude Sonnet 5.5 | Anthropic | 1.0M tokens | n/a | Proprietary | 85/100 | Sep 28, 2026 |
| Claude Haiku 5.5 | Anthropic | 1.0M tokens | n/a | Proprietary | 70/100 | Oct 7, 2026 |
| GPT-6.1 Sol | OpenAI | 1.1M tokens | n/a | Proprietary | 40/100 | Sep 29, 2026 |
| GPT-6.1 Sol Pro | OpenAI | 1.1M tokens | n/a | Proprietary | 40/100 | Sep 29, 2026 |
| Kimi K3 | Moonshot AI | n/a | n/a | Kimi K3 License | n/a (AA Index 44) | Jul 16, 2026 |
Scores above come from LM Market Cap’s composite methodology where listed, with Kimi K3 shown on the Artificial Analysis Intelligence Index since it was not in LM Market Cap’s tracked coding set at the time of writing. Note that GPT-6.1 Sol Pro is, per its own listing, “the same underlying model as GPT-6.1 Sol, served with reasoning.mode set to pro for higher-quality responses on complex tasks,” which is why the two share an identical score despite different marketing names. OpenAI’s own documentation flags that pro mode consumes substantially more tokens per response, which matters more for cost than for the headline score.
How the Scoring Works: Inside the LM Market Cap Composite Score
Because this entire comparison leans on a tracker most readers have never heard of, it is worth spelling out exactly what that composite score measures. According to LM Market Cap’s published methodology, every model it tracks receives a score from 0 to 100 built from three weighted components: benchmark performance makes up 90 percent of the score, aggregated and normalized across Arena Elo ratings, MMLU, GPQA, HumanEval, SWE-bench, MATH, GSM8K, IFEval, and more than ten additional standardized evaluations pulled from multiple independent sources. Capabilities, meaning feature breadth such as vision, function calling, and reasoning support, contribute 5 percent as a tiebreaker. Context window size relative to the field contributes the remaining 5 percent, also as a tiebreaker. The site states the data refreshes hourly via automated pulls from more than 68 providers and currently tracks over 466 models, of which 167 are open source and 29 are free.
That transparency matters because a single-number score can otherwise look arbitrary. Knowing that 90 percent of the Qwen3.8 Max Prime, GLM 5.3 Prime, and Step 5 Preview scores trace back to standardized coding and reasoning benchmarks, rather than vibes or marketing claims, gives the 40-vs-75-vs-79 spread between these three releases some actual weight. It also explains why GPT-6.1 Sol and GPT-6.1 Sol Pro land at an identical 40, since LM Market Cap is scoring benchmark outcomes, and OpenAI’s own description confirms the two are the same base model with a different reasoning-effort setting.
Benchmark Results from Three Independent Trackers
No single benchmark tells the whole story here, so this section lines up results from LM Market Cap, Artificial Analysis, and the aggregated table on Wikipedia’s generative-AI comparison page, which itself pulls from Arena.ai, the Epoch Capability Index, and the Vals Index.
| Model | LM Market Cap Score | AA Intelligence Index | Arena.ai Text Rating | Epoch Capability Index | Vals Index |
|---|---|---|---|---|---|
| Qwen3.8 Max (0902) | 40 | 45 | 1,483 | 156.4 | 48.3 |
| GLM-5.3 | 79 | 45 | 1,478 | 155.6 | 53.5 |
| Step 5 Preview | 75 | 44 | 1,447 | n/a | n/a |
| Kimi K3 | n/a | 44 | 1,488 | 157.4 | 50.3 |
| Grok 4.7 | n/a | 46 | 1,443 | 153.5 | 55.0 (approx.) |
| Mistral Large 4 Preview | 72 | 38 | n/a | n/a | n/a |
| GPT-6.1 Sol | 40 | 52 | 1,484 | 166.1 | 61.2 |
| GPT-6 Astra | n/a | 53 | 1,475 | 166.4 | 63.1 |
| Claude Sonnet 5.5 | 85 | 56 | 1,476 | 165.0 | 67.0 |
| Claude Opus 5.5 | 95 | 58 | 1,507 | 167.3 | 67.0 |
The interesting read here is not that any Chinese AI model has overtaken the US frontier tier. On the Artificial Analysis Intelligence Index, which Wikipedia uses as its primary sorting column, Claude Opus 5.5 leads at 58, Claude Sonnet 5.5 follows at 56, and GPT-6.1 Sol sits at 52. Qwen3.8 Max and GLM-5.3 tie at 45, with Step 5 Preview and Kimi K3 just behind at 44. That puts the best of the new Chinese AI model trio roughly 7 points behind GPT-6.1 Sol and 13 points behind Claude Opus 5.5 on this particular index, a real but no longer enormous gap. Where the story flips entirely is the LM Market Cap composite score, which weighs a broader basket of coding-specific benchmarks more heavily: there, GLM-5.3 and GLM 5.3 Prime (79) actually outscore GPT-6.1 Sol and GPT-6.1 Sol Pro (40), while Qwen3.8 Max Prime lands at the same 40. Different benchmark baskets produce different rankings, which is exactly why this comparison uses three sources instead of one.
Pricing Breakdown: Cost Per Million Tokens
Pricing is where the three new releases diverge most sharply, and where the “Chinese AI models are always cheaper” assumption breaks down. Qwen3.8 Max Prime is actually the most expensive model in this entire comparison on a per-token basis, more expensive than Claude Sonnet 5.5 and in the same range as Claude Opus 5.5’s output pricing.
| Model | Input ($/M tokens) | Output ($/M tokens) | AA Cost Per Task |
|---|---|---|---|
| Step 5 Preview | $1.00 | $2.70 | $1.03 |
| GLM-5.3 (base) | $0.08 | $4.80 | $2.01 |
| GLM 5.3 Prime | $2.80 | $8.80 | n/a (base used for AA) |
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.21 |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.72 |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $5.46 |
| Qwen3.8 Max Prime | $4.00 | $12.00 | $5.41 (base Max) |
| Claude Opus 5.5 | $4.00 | $20.00 | $5.98 |
Step 5 Preview undercuts everything else in this table on raw per-token pricing, at $1.00 input and $2.70 output per million tokens, and it also posts the lowest Artificial Analysis cost-per-task figure at $1.03. GLM-5.3’s base tier is even cheaper on input tokens at just $0.08 per million, though its $4.80 output price and the Prime variant’s full $2.80/$8.80 pricing put it solidly in the middle of the pack once a full conversation’s worth of output is factored in. Qwen3.8 Max Prime, meanwhile, charges $4.00 input and $12.00 output per million tokens, matching or exceeding Claude Opus 5.5’s input price while trailing it badly on the LM Market Cap composite score. For any team choosing between these three purely on cost efficiency, Step 5 Preview is the clear value play, GLM 5.3 sits in the middle depending on which tier is used, and Qwen3.8 Max Prime is a premium-priced option that, based on current scoring data, does not currently deliver premium-tier results to match.
Speed and Cost-Per-Task: The Hidden Gap
Per-token pricing only tells half the story, because models differ wildly in how many tokens they burn to finish a task, and in how fast they generate those tokens. Artificial Analysis’s median tokens-per-second figures expose a gap that raw pricing tables hide entirely. Step 5 Preview generates at a median of 87 tokens per second, the fastest of the three new releases. GLM-5.3 follows at 83 tokens per second. Qwen3.8 Max trails badly at just 36 tokens per second, less than half the speed of its two rivals, which compounds its higher per-token price into an even larger real-world cost and latency disadvantage for any agentic workflow that depends on fast turnaround.
For comparison, Claude Sonnet 5.5 generates at 141 tokens per second, nearly four times Qwen3.8 Max’s throughput, while Claude Opus 5.5 runs at 96 tokens per second despite being the most capable model in this entire comparison set. GPT-6.1 Sol sits at a more modest 56 tokens per second. The takeaway for anyone building a latency-sensitive application, such as a live coding assistant or a conversational agent, is that Qwen3.8 Max Prime’s combination of premium pricing and below-average generation speed makes it the weakest choice in this comparison for interactive use cases, regardless of its branding as Alibaba’s flagship tier.
Licensing: Which Model Is Actually Open
Open-weight access is often assumed to be a given with Chinese AI model releases, but that assumption does not hold for this particular trio. Wikipedia’s comparison table lists GLM-5.3 under an MIT license, meaning Zhipu AI has published the base model’s weights for download and local deployment, a meaningful differentiator for teams that need to self-host or fine-tune. Both Qwen3.8 Max and Step 5 Preview, by contrast, are listed as proprietary, meaning access runs exclusively through hosted APIs rather than downloadable weights, at least for the flagship Max and Preview tiers covered in this comparison. That licensing split cuts directly against the common shorthand that Chinese labs default to open release and US labs default to closed release. Here, it is an American rival, Meta’s Muse Spark 1.3, that remains proprietary per Wikipedia’s table, while Zhipu AI, a Chinese lab, is the one shipping MIT-licensed weights for its flagship coding model. Alibaba does maintain a separate, smaller Qwen3.8 27B model with broader availability on model hubs, but that SKU scores dramatically lower (20 to 28 on the Artificial Analysis Intelligence Index) than the proprietary Max Prime tier benchmarked throughout this piece, so it is not a substitute for teams specifically evaluating Qwen3.8 Max.
Where to Access Each Model Today
All three models are live and bookable through multiple channels as of this writing, which is itself useful signal, since a model stuck behind a waitlist is a different proposition from one with five working integration paths.
- OpenRouter lists playground access for all three models under the identifiers stepfun/step-5-preview, z-ai/glm-5.3-prime, and qwen/qwen3.8-max-prime, making it the fastest way to test all three side by side without separate vendor accounts.
- Alibaba Cloud Model Studio serves Qwen3.8 Max and Qwen3.8 Max Prime directly as Alibaba’s own hosted endpoint.
- StepFun’s own API and consumer products remain the primary route to Step 5 Preview for teams building directly against StepFun’s infrastructure rather than through an aggregator.
- Zhipu AI (Z.ai) hosts both the base GLM-5.3 and the GLM 5.3 Prime variant through its own developer platform, alongside the MIT-licensed weights for teams that want to self-host the base model instead.
- Third-party inference platforms have also picked up individual models faster than expected; Artificial Analysis’s own changelog logs new provider evaluations for comparable releases, including additional hosts for DeepSeek and Mistral models within the same week, which suggests the broader open-weight and high-throughput model hosting market is moving quickly enough that availability for any single model can change within days of this being published.
5 Real-World Use Cases for Each Model
Given the pricing and benchmark spread above, these three releases do not map onto a single “best overall” use case. Instead, each one fits a different slice of a typical engineering or product workflow.
- High-volume batch coding tasks: Step 5 Preview’s $1.00/$2.70 pricing and 87 tokens-per-second speed make it the strongest fit for large-scale automated code review, test generation, or documentation jobs where thousands of calls run unattended and cost compounds quickly.
- Self-hosted or air-gapped deployments: GLM-5.3’s MIT license is the only genuinely open-weight option of the three, making it the default choice for regulated environments, on-premises deployments, or teams that need to fine-tune on proprietary data without sending it to a third-party API.
- Multimodal agentic tasks involving video: Qwen3.8 Max Prime’s support for text, image, and video input is the one capability none of its two rivals in this piece advertise, making it a reasonable pick specifically for workflows that need to reason over video content, despite its weaker coding-benchmark showing.
- Latency-sensitive interactive assistants: Neither GLM 5.3 Prime’s 83 tokens/second nor Step 5 Preview’s 87 tokens/second match Claude Sonnet 5.5’s 141 tokens/second, but both comfortably beat Qwen3.8 Max’s 36 tokens/second, so teams building chat-style tools on a China-hosted stack should default to GLM 5.3 Prime or Step 5 Preview over Qwen3.8 Max Prime.
- Budget-constrained prototyping: Base GLM-5.3’s $0.08 input price is the cheapest entry point in this entire comparison, making it the natural choice for early-stage experimentation before committing to a production pricing tier.
How They Compare to GPT-6.1 Sol and Claude Opus 5.5
The honest framing, based on the data gathered above, is that this new Chinese AI model trio has narrowed the gap to the US frontier tier on some measures while widening it on others. On the Artificial Analysis Intelligence Index, the best of the three (Qwen3.8 Max and GLM-5.3, tied at 45) sits 7 points behind GPT-6.1 Sol’s 52 and 13 points behind Claude Opus 5.5’s 58. That is a real, measurable gap, not closed, but also not the multi-generation chasm that existed a year or two earlier in the broader AI model race.
On LM Market Cap’s composite score, which leans more heavily on coding-specific benchmarks, the picture flips: GLM-5.3 and GLM 5.3 Prime’s score of 79 actually beats GPT-6.1 Sol’s 40, and Step 5 Preview’s 75 does too, though neither comes close to Claude Opus 5.5’s 95 or Claude Sonnet 5.5’s 85. Qwen3.8 Max Prime, at 40, ties GPT-6.1 Sol on that scale rather than beating it. Readers should treat “beats GPT-6.1 Sol” and “trails Claude Opus 5.5 by double digits” as both true simultaneously, since they are measuring different things: LM Market Cap’s basket is weighted toward coding benchmarks where several Chinese AI model releases have specifically optimized, while the Artificial Analysis Intelligence Index spans a broader set of reasoning and knowledge tasks where the US frontier tier still leads more consistently. For deeper context on how Claude Opus 5.5 stacks up against OpenAI’s other recent release, see the prior comparison of Claude Opus 5.5 and GPT-6 Astra, and for GPT-6.1 Sol’s own head-to-head against Google’s newest model, the GPT-6.1 Sol vs. Gemini 4 Argon breakdown covers that matchup in detail.
Migration Guide: Moving Between the Three Models
Teams already running production workloads on one of these models, or on a comparable release like DeepSeek V4.1 Flash or Mistral Large 4, can move between Qwen3.8 Max Prime, GLM 5.3 Prime, and Step 5 Preview with relatively little code churn if they are already routing through OpenRouter or a similar aggregator, since all three expose standard chat-completion style endpoints.
- Audit current token usage patterns first. Since Qwen3.8 Max Prime charges $12.00 per million output tokens against Step 5 Preview’s $2.70, a migration away from Qwen3.8 Max Prime toward Step 5 Preview could cut output costs by more than 75 percent for equivalent workloads, but only if output quality holds up for the specific task.
- Check context window ceilings against real usage. All three models advertise roughly 1.0-million-token context windows, but maximum output differs substantially: 131.1K for Qwen3.8 Max Prime, 131K for GLM 5.3 Prime, and just 64K for Step 5 Preview. Workloads that generate long single responses, such as full-file rewrites, need to confirm Step 5 Preview’s lower output ceiling will not truncate results.
- Swap the model identifier in the API call. Through OpenRouter, this is typically a single string change, for example from qwen/qwen3.8-max-prime to z-ai/glm-5.3-prime or stepfun/step-5-preview, without touching the surrounding request structure.
- Re-run function-calling and JSON-mode tests explicitly. All three support these features per their model cards, but exact schema adherence and tool-call formatting can differ enough between vendors to break brittle parsing logic, so this step should not be skipped even when the feature list looks identical on paper.
- Benchmark real latency, not just advertised speed. Artificial Analysis’s reported median speeds (87 tokens/second for Step 5 Preview, 83 for GLM-5.3, 36 for Qwen3.8 Max) are averages across their test suite and can vary by provider and region, so teams should run their own timed requests against the specific hosting endpoint they plan to use in production.
- If self-hosting is the goal, only GLM-5.3’s base tier currently offers that path, given its MIT license; migrating to a self-hosted GLM-5.3 deployment requires separate infrastructure planning that a simple API-key swap does not cover.
// Example: swapping model identifiers via an OpenRouter-style endpoint
const response = await fetch("https://openrouter.ai/api/v1/chat/completions", {
method: "POST",
headers: {
"Authorization": `Bearer ${process.env.OPENROUTER_API_KEY}`,
"Content-Type": "application/json"
},
body: JSON.stringify({
model: "stepfun/step-5-preview", // swap to "z-ai/glm-5.3-prime" or "qwen/qwen3.8-max-prime"
messages: [{ role: "user", content: "Refactor this function for readability." }],
max_tokens: 4096
})
});Pros and Cons of Each Model
Qwen3.8 Max Prime
Pros: Only model in this comparison with native video input support; backed by Alibaba Cloud’s broader enterprise infrastructure; 131.1K max output.
Cons: Most expensive per-token pricing in the comparison ($4.00/$12.00); lowest LM Market Cap score among the three new releases (40/100); slowest generation speed at 36 tokens/second; proprietary, no self-hosting path.
GLM 5.3 Prime (and base GLM-5.3)
Pros: Highest LM Market Cap score of the three (79/100); base model is MIT-licensed and genuinely open-weight; cheapest input pricing in the entire comparison at the base tier ($0.08/M); solid 83 tokens/second throughput.
Cons: Prime tier’s $2.80/$8.80 pricing is a meaningful step up from the base model; output ceiling and context window are not class-leading against the newest US models; text-only input, no native vision/video intake for the base model per its feature list.
Step 5 Preview
Pros: Cheapest per-token pricing in the comparison ($1.00/$2.70); fastest generation speed at 87 tokens/second; strong 75/100 LM Market Cap score; supports vision alongside text.
Cons: Lowest maximum output ceiling of the three at 64K tokens; newest and least battle-tested in production, having only shipped October 8, 2026; proprietary with no confirmed open-weight path at the time of writing.
The Broader Open-Weight Landscape Around This Release Window
Qwen3.8 Max Prime, GLM 5.3 Prime, and Step 5 Preview did not launch into a quiet market. The same late-September-to-October window produced a cluster of smaller, cheaper releases worth knowing about even if they sit outside this comparison’s main three. InclusionAI’s Ling 3.1 Flash, a 560-billion-parameter model, appeared in September 2026, while a companion release, Ling 3.0 Flash Fin, shipped as a 124-billion-parameter model with a 262,000-token context window in the same period, according to LM Market Cap’s tracking data. IBM’s Granite 4.2 8B landed on August 31, 2026, with a far smaller 131,000-token context window and a composite score of 54, a reminder that not every new release in this window is chasing frontier-scale benchmarks. Inception’s Mercury 2.5, released August 31, 2026, takes a different architectural approach entirely, running as a diffusion language model that reportedly generates text at roughly 1,107 tokens per second on standard GPUs, far outpacing every autoregressive model in this comparison on raw throughput, though at a different point on the quality-versus-speed curve that standard benchmark suites do not capture cleanly.
That context matters for anyone using this piece to shop for a model rather than just to read about three specific releases. Qwen3.8 Max Prime, GLM 5.3 Prime, and Step 5 Preview all target the same agentic-coding, large-context niche, but they are competing in a field that also includes much smaller, cheaper, and architecturally different options from InclusionAI, IBM, and Inception shipping in the same few weeks. A team that does not specifically need a 1-million-token context window or top-quartile coding benchmarks may find one of these smaller releases, or a lighter-weight model like Granite 4.2 8B, a better fit for cost reasons alone, even though none of them appear in the head-to-head tables above because they are not direct competitors on specs.
The Verdict: Which Chinese AI Model Should You Actually Use
Based on the pricing, speed, and benchmark data gathered across LM Market Cap, Artificial Analysis, and Wikipedia’s aggregated comparison table, GLM 5.3 (either the MIT-licensed base model or the faster Prime variant) comes out ahead for most general-purpose use. It posts the highest composite score of the three (79/100), offers the only genuine open-weight option, and prices its base tier lower than either rival. Step 5 Preview is the right pick specifically when raw cost-per-token and generation speed matter more than maximum output length, since its $1.00/$2.70 pricing and 87 tokens/second speed beat both competitors on throughput economics, even though its 64K output ceiling and newcomer status (released October 8, 2026) warrant more caution before committing to it for mission-critical workloads. Qwen3.8 Max Prime is the hardest of the three to recommend on the numbers alone: it costs more than Claude Sonnet 5.5, scores the same as GPT-6.1 Sol on LM Market Cap’s coding-weighted scale, and runs at less than half the speed of its two Chinese AI model rivals. Its one clear differentiator, native video input, is worth paying for only if that specific capability is part of the actual product requirement, not as a general-purpose coding or reasoning upgrade.
Zoomed out, the broader trend holds regardless of which of these three models a given team picks: Chinese AI model releases are no longer trailing the US frontier tier by a wide, generation-defining margin on every metric. The Artificial Analysis Intelligence Index puts the best of this trio within single digits of GPT-6.1 Sol, and LM Market Cap’s coding-weighted score has GLM-5.3 beating it outright. But “closed the gap” is not the same as “interchangeable,” and the near-4x price spread and more than 2x speed spread between Qwen3.8 Max Prime and its two rivals shows that picking the wrong model from this group, purely on the basis of country of origin or branding, is still an easy and costly mistake to make.
Frequently Asked Questions
Is GLM 5.3 open source?
The base GLM-5.3 model carries an MIT license according to Wikipedia’s comparison table, making it genuinely open-weight and available for self-hosting. The faster GLM 5.3 Prime variant is served as a separate hosted SKU at a higher price, inheriting the base model’s capabilities with 1.5x to 2x the output throughput.
How much does Step 5 Preview cost per million tokens?
Step 5 Preview prices at $1.00 per million input tokens and $2.70 per million output tokens, according to LM Market Cap, making it the cheapest per-token option among the three new releases covered in this comparison.
What is the difference between Qwen3.8 Max and Qwen3.8 Max Prime?
Qwen3.8 Max Prime is a higher-throughput variant of the base Qwen3.8 Max, served as a separate SKU at a higher price point, per its LM Market Cap listing. Both share the same underlying capabilities, including text, image, and video input, but Prime is optimized for faster inference at a premium price.
Which Chinese AI model scores highest on independent benchmarks?
GLM-5.3 and GLM 5.3 Prime post the highest LM Market Cap composite score among the three (79/100), and GLM-5.3 ties Qwen3.8 Max for the highest Artificial Analysis Intelligence Index score (45) among this specific trio, per Wikipedia’s tracking table.
How do these three Chinese AI models compare to GPT-6.1 Sol and Claude Opus 5.5?
On the Artificial Analysis Intelligence Index, the strongest of the three (tied at 45) trails GPT-6.1 Sol’s 52 by 7 points and Claude Opus 5.5’s 58 by 13 points. On LM Market Cap’s coding-weighted composite score, GLM-5.3 and Step 5 Preview both outscore GPT-6.1 Sol’s 40, though none of the three approach Claude Opus 5.5’s 95.
Can I access these models outside China?
Yes. All three are listed on OpenRouter’s playground under the identifiers stepfun/step-5-preview, z-ai/glm-5.3-prime, and qwen/qwen3.8-max-prime, in addition to their respective vendor platforms (StepFun’s own API, Zhipu AI’s developer platform, and Alibaba Cloud Model Studio).
What does the LM Market Cap score actually measure?
It is a composite score from 0 to 100, weighted 90 percent toward normalized benchmark performance (covering Arena Elo, MMLU, GPQA, HumanEval, SWE-bench, MATH, GSM8K, IFEval, and more), with capabilities and context window each contributing 5 percent as tiebreakers, according to the site’s published methodology.
Which model is cheapest for high-volume coding tasks?
Step 5 Preview, at $1.00 input and $2.70 output per million tokens with an $1.03 Artificial Analysis cost-per-task figure, is the cheapest of the three for high-volume use, though base GLM-5.3’s $0.08 input price is lower still if output length per task is kept short.
Related Coverage
- Claude vs Copilot 2026: $20 vs $12.99 Price Gap
- Nano Banana vs GPT Image API: 13-Step AI Benchmark [2026]
- DeepSeek V4.1 Flash API Setup: 13 Steps, 1M Tokens [2026]
- Cloudflare Clef-omni: 30B AI Model Spans 4 Formats [2026]
- Bill Gates Warns AI Will Erase Jobs, Urges US-Led Rules [2026]
