GLM-5.2 Nearly Matches Opus 4.8, Kimi K2.7-Code Improves at the Same Price, and Gemini Image Goes GA

I. Update Overview

In June, we launched 4 new models, concentrated across two main capability tracks, along with retirement schedules for several legacy models that require advance migration planning.

The two capability tracks are distributed as follows:

The coding track is the busiest. Moonshot released Kimi K2.7-Code and its high-speed variant in one move, while Zhipu’s GLM-5.2 also focuses on long-horizon autonomous coding. The key theme for this track is: “pricing is largely unchanged, while capability has clearly moved upward.”

The image track has reached a “GA” milestone. Google’s Gemini 3 Pro Image, also widely known as Nano Banana Pro, has moved from preview to general availability (GA), replacing the preview version that will soon be retired. This is suitable for teams that want to put image generation into production workflows.

Below is the daily launch list, followed by deep dives into the key models.

II. Daily Launch List

June 17

GLM-5.2 (Zhipu AI): the latest language model in the GLM series, focused on long-horizon autonomous coding.

Kimi K2.7-Code Series (Moonshot AI): coding-specialized models, including the standard version and the highspeed variant.

Gemini 3 Pro Image / Nano Banana Pro (Google): the GA version of the image generation model, replacing the preview version.

Synchronized note: this June batch is mainly about new model launches. In addition, three groups of legacy models from Kimi K2, GLM, and Gemini have entered retirement or deprecation schedules. Please arrange migrations accordingly. See Section VI for details.

III. Deep Dive into Key Models

3.1 GLM-5.2 (Zhipu AI) — Near First-Tier Coding Scores at the Same Price Point

Launch date: launched on AF on June 17.

Provider pricing: token-based billing; $1.4 / 1M input tokens, $4.4 / 1M output tokens, and $0.26 / 1M cache read. Pricing remains the same as the previous-generation GLM-5.1.

Primary scenarios: long-horizon autonomous coding and engineering tasks, agent workflows, 1 million-token context, and up to 131K output tokens.

Horizontal comparison

Viewed within Zhipu’s own product line, GLM-5.2 pushes coding and agent capabilities up another level while keeping pricing unchanged from GLM-5.1. It is a typical “more capability at the same price” iteration.

Across vendors, in benchmarks released by Zhipu on June 16, GLM-5.2 scored 62.1 on SWE-bench Pro, exceeding GPT-5.5 at 58.6; it scored 74.4% on FrontierSWE, above GPT-5.5’s 72.6% and less than one point behind Claude Opus 4.8’s 75.1%; and it scored 77.0 on MCP-Atlas, also ahead of GPT-5.5’s 75.3 [1][5]. In pricing, GPT-5.5 is approximately $10 / $30, Claude Opus 4.8 standard tier is $5 / $25, while GLM-5.2 is priced at $1.4 / $4.4, only a fraction of the former two.

One of the most visible signals this month came from security. In late June, security vendor Semgrep published cybersecurity benchmarks showing that GLM-5.2 achieved 39% F1 on IDOR (insecure direct object reference) vulnerability detection without any additional scaffolding, outperforming Claude Code running in the same unscaffolded setting at 32%, with an estimated cost of about $0.17 per vulnerability found. On agentic offensive-defense tasks such as CyBT-CTF, it is currently the strongest open-weight model, with a solve rate on par with Anthropic Opus 4.7/4.8 [9]. Two points need to be made clear: first, this is a “model-to-model, no dedicated pipeline” comparison, and GLM-5.2 still trails Semgrep’s own multimodal security pipeline at 53–61% F1; second, on harder general reasoning tasks such as SWE-Marathon, GLM-5.2 still significantly trails Opus 4.8. In other words, its price-performance advantage is most prominent in “high-volume, batchable” scenarios such as security scanning.

AF use case: A SaaS engineering team uses AF’s project management capability to classify engineering tasks. High-volume agent tasks such as regression test fixes, batch refactoring, and dependency upgrades are assigned to GLM-5.2, while the most difficult core module reviews are assigned to Opus 4.8. By matching models to task types and assigning separate budgets and quotas, the team reduces the cost of high-volume stages to a fraction of the original level under the same token budget, while noticeably improving overall iteration speed.

3.2 Kimi K2.7-Code Family (Moonshot AI) — Same Price, Higher Scores, and Better Token Efficiency

This month, Moonshot released two models in the same family. We discuss them together:

kimi-k2.7-code: $0.95 / 1M input tokens, $4 / 1M output tokens, and $0.19 / 1M cache read.

kimi-k2.7-code-highspeed: $1.9 / 1M input tokens, $8 / 1M output tokens, and $0.38 / 1M cache read. It is priced at twice the standard version in exchange for lower latency, making it suitable for interactive coding scenarios that are sensitive to response speed.

Launch date: launched on AF on June 17.

Architecture and billing model: MoE architecture with 1T total parameters, approximately 32B active parameters per token, 256K context window; token-based billing.

Primary scenarios: long-horizon software engineering — planning, code modification, tool use, and multi-step debugging — designed for coding agents rather than general chat.

Horizontal comparison

Within Moonshot’s own model line, the most important point about K2.7-Code is that the standard version is priced exactly the same as the previous-generation Kimi K2.6 official price ($0.95 / $4), while capability clearly improves. Moonshot reports that, compared with K2.6, Kimi K2.7-Code improves by 21.8% on Kimi Code Bench v2, 11.0% on Program Bench, and 31.5% on MLS Bench Lite, while reducing reasoning token usage by about 30% [3]. In other words, for the same coding task, the unit price has not increased, and the number of tokens required to complete the task is lower, so the actual bill decreases.

Across vendors, compared with GLM-5.2 ($1.4 / $4.4) and Claude Opus 4.8 ($5 / $25), both positioned for coding, Kimi K2.7-Code’s $0.95 / $4 makes it one of the lowest-cost open-weight options in this tier. It is suitable as a high-frequency, high-volume coding agent model.

（sheets)

AF use case: An internal developer platform turns “auto-fix lint issues, add unit tests, and generate PR descriptions” into always-on agents. The standard kimi-k2.7-code is used by default to control costs; when engineers are pair-programming inside the IDE and need second-level responses, the workflow temporarily switches to the highspeed variant for lower latency. Since the two models are in the same family and use the same interface, switching requires no code changes.

3.3 Gemini 3 Pro Image / Nano Banana Pro (Google) — From Preview to GA, Ready for Production

Launch date: launched on AF on June 17 as the GA version, replacing the soon-to-be-retired gemini-3-pro-image-preview.

Provider pricing: token-based billing; $2 / 1M input tokens, $12 / 1M text output tokens, and $120 / 1M image output tokens. Converted to per-image cost: about $0.039 per image for 1024×1024 and below, about $0.134 for the 1K–2K tier, and about $0.24 for 4K (4096×4096). Batch processing can cut these costs in half, at the cost of up to 24 hours of processing latency [4].

Primary scenarios: high-quality image generation and editing, especially text rendering inside images, 4K output, and instruction-based fine editing.

Horizontal comparison

Within Google’s own model line, the biggest significance of the GA version compared with preview is stability and commercially supportable SLA. The preview version has already been scheduled for retirement on July 17, so teams that want to use it in production workflows should move directly to this GA version. Its sibling model gemini-2.5-flash-image follows a cheaper and faster lightweight path, but has been scheduled for deprecation on October 2.

Across vendors, ByteDance’s Seedream 4.5 starts at around $0.03 per image, clearly below Nano Banana Pro’s $0.134 per image in the 1K–2K tier [6]. Beyond price, Nano Banana Pro’s strengths lie in text rendering inside images, 4K resolution, and complex instruction-based editing. For bulk image generation that prioritizes the lowest unit price and does not require demanding text rendering, models such as Seedream are more cost-effective. A more stable setup is to keep both options available and select the model according to the image task.

AF use case: An e-commerce team uses Nano Banana Pro to batch-generate hero images and product detail images with Chinese promotional text. The key reason for choosing it is that text inside the image remains undistorted and can be refined according to brand guidelines. For pure background images or atmospheric visuals without text, the team uses cheaper image models for volume.

IV. Pricing and Billing Comparison

The table below uses Anthropic models as comparison anchors and lists provider pricing for key models in this update. All are token-based and shown in $/1M.

（sheets)

Additional billing notes:

For image models, per-image cost is more intuitive after token conversion. Gemini 3 Pro Image is approximately $0.039 per image (≤1K), $0.134 per image (1K–2K), and $0.24 per image (4K). Batch processing cuts the cost in half.

Cache hits are most cost-effective for high-frequency agent tasks with long fixed prefixes. GLM-5.2 and Kimi K2.7-Code both provide very low cache read pricing. Used together with fixed system prompts within projects, this can reduce bills by another tier.

Billing is subject to each model provider’s official billing model and promotion policies. AF pricing is aligned with the original provider.

V. Enterprise Model Selection Recommendations

High-frequency coding agents: use kimi-k2.7-code by default, as it has the lowest unit price and high token efficiency. Switch to the highspeed variant when latency is critical. On AF, you can create a “coding agent” project according to your own rules, place both models into the same invocation order, and define which scenarios use the standard version and which use the highspeed version.

Engineering tasks that prioritize quality while controlling cost: use GLM-5.2 for the bulk of work and assign the most difficult core reviews to Claude Opus 4.8. On AF, you can freely switch APIs across these providers, customize model invocation order, and use project management to assign budgets, quotas, and permissions to different models — reducing costs for high-volume tasks while preserving quality for critical tasks.

Production image generation: use the GA version of Gemini 3 Pro Image for high-value assets requiring text rendering, 4K output, and fine editing. Pair it with cheaper image models for bulk materials such as pure background and atmosphere images. On AF, image tasks and text tasks can be managed under separate project budgets to prevent overspending on one line from affecting the entire account.

The overall idea can be summarized in one sentence: keep model selection control in the hands of you and your agents. Create projects according to your own rules, freely switch APIs, customize invocation order, and use project management to allocate budgets, quotas, and permissions.

VI. Retirement and Migration Notices

This issue includes retirement and deprecation schedules for three groups of legacy models. Please arrange migrations in advance to avoid service interruption.

6.1 Kimi K2 Series Retirement (Effective May 25)

Moonshot AI officially retired the Kimi K2 series on May 25, 2026, and no longer maintains or supports it.

（sheets)

In addition, the three legacy moonshot-v1-*-vision-preview visual preview models (8k / 32k / 128k) have been marked as deprecated. Please migrate to the Kimi K2 series as soon as possible.

6.2 GLM Model Deprecation

Zhipu AI has announced that the following GLM models are entering the deprecation process:

（sheets)

Please plan the migration as soon as possible. Contact us directly if you need technical assistance.

6.3 Gemini Preview Model Deprecation Schedule

Google plans to clean up and deprecate certain Gemini Preview models according to the following schedule:

（sheets)

We recommend migrating to replacement models such as gemini-3.5-flash and gemini-3.1-pro-preview. For gemini-3-pro-image-preview, the replacement is the GA version of Gemini 3 Pro Image launched in this update. You can switch directly.

The main theme this month is clear: coding models are collectively moving toward “stronger at the same price, with fewer tokens,” while image models are moving from preview to production-ready GA versions. The most valuable way to use this batch is to assign high-volume work to cheaper models, reserve high-end models for critical work, and manage budgets and permissions through projects.

If you need help organizing a migration list or building a model combination based on your task types, contact the AF team at any time.

References

[1] VentureBeat — Z.ai's open-weights GLM-5.2 beats GPT-5.5: https://venturebeat.com/technology/z-ais-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-long-horizon-coding-benchmarks-for-1-6th-the-cost

[2] CloudPrice — GLM-5.2 pricing & specs: https://cloudprice.net/models/zhipu-glm-5-2

[3] MarkTechPost — Moonshot AI Releases Kimi K2.7-Code: https://www.marktechpost.com/2026/06/12/moonshot-ai-releases-kimi-k2-7-code-a-coding-model-reporting-21-8-on-kimi-code-bench-v2-over-k2-6/

[4] PricePerToken — Nano Banana Pro (Gemini 3 Pro Image) pricing: https://pricepertoken.com/pricing-page/model/google-gemini-3-pro-image-preview

[5] EdenAI — GLM-5.2 Benchmark vs GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro: https://www.edenai.co/post/glm-5-2-benchmark-vs-gpt-5-5-claude-opus-4-8-and-gemini-3-1-pro

[6] OpenRouter — Seedream 4.5 API pricing & benchmarks: https://openrouter.ai/bytedance-seed/seedream-4.5

[7] Anthropic / Claude API pricing: https://platform.claude.com/docs/en/about-claude/pricing

[8] TokenMix — Kimi API Pricing 2026 (K2.6 $0.95, etc.): https://tokenmix.ai/blog/kimi-k2-api-pricing

[9] Semgrep — GLM 5.2 beats Claude in our Cyber Benchmarks (IDOR F1 benchmark): https://semgrep.dev/blog/2026/we-have-mythos-at-home-glm-52-beats-claude-in-our-cyber-benchmarks/