AI Coding Agents Report: June 2026

01 July 2026 · Version history

Executive Summary

June 2026 was defined by government intervention in frontier-model access, a painful and uneven landing for metered billing, and a step-change in open-weight quality. On June 12, a US export-control directive forced Anthropic to suspend Fable 5 and Mythos 5 for all customers, including foreign-national employees. Fable 5 was cleared to return July 1 while Mythos 5 stays suspended with no timeline. Two weeks later OpenAI previewed GPT-5.6 (Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per MTok) but, at the government's request, staggered the release and vetted who gets preview access. Anthropic responded by proposing an industry jailbreak-severity framework with Amazon, Microsoft, and Google. Frontier access is now non-guaranteed.

The subscription-to-metered transition landed unevenly. Copilot AI Credits went live June 1 as predicted, but users report burning a month of credits in days (nativeit). Anthropic paused its Agent SDK credit change on June 15 with no replacement date. Augment retired credits and individual tiers entirely, moving to a $100/mo Business flat plan plus a 40% pass-through fee. Meanwhile SpaceX signed a definitive $60B deal to acquire Cursor on June 16, and Elon Musk teased Grok 4.5 "with Cursor data added in."

The open side closed the gap. Zhipu's MIT-licensed GLM-5.2 reached near-Opus-4.8 coding quality at roughly one-sixth the price. May's predictions split: AI Credits live (true), OpenAI promos expired (true), Gemini CLI shutdown (true), but Anthropic SDK credits (false, paused), Gemini 3.5 Pro (false, slipped to July), Alibaba's promo (false, extended to July 22), Zhipu's multiplier (false, extended to September), and Tabnine v6.3 (false, silently slipped).

Model & Capability Developments

June delivered the strongest open-weight coding models this project has tracked. Anthropic's Sonnet 5 launched June 30 at introductory $2/$10 per MTok (rising to $3/$15 after Aug 31), positioned close to Opus 4.8 quality. But its updated tokenizer produces roughly 30% more tokens (1.42x English, 1.28x Python), which Simon Willison calls "effectively a 30% price increase" once the intro window expires. Zhipu's GLM-5.2 (MIT-licensed, June 16) is the bigger step: a ~744B/40B MoE that Artificial Analysis ranks first among open models at Intelligence Index 51, level with GPT-5.5 on agentic GDPval and between Opus 4.7 and 4.8 on long-horizon coding, at $1.40/$4.4 per MTok. Moonshot's open-weight Kimi K2.7 Code ($0.95/$4.00) scores 62.0 on Kimi Code Bench v2 (up from 50.9) but trails GPT-5.5 (69.0) and Opus 4.8 (67.4) on every row.

On the frontier-but-restricted side, GPT-5.6 adds a new ultra subagent mode and max reasoning, claiming a Terminal-Bench 2.1 state of the art, but scores are published as charts only. Anthropic's Mythos-class Fable 5 and Mythos 5 ($10/$50 per MTok) were suspended June 12; Fable 5 returns July 1, Mythos 5 does not.

In agentic paradigms, Cognition's Devin Fusion runs a frontier model plus a cheaper sidekick in parallel for 35% lower cost at matched quality, and router proxies like the Weave Router emerged to escape frontier token costs. On inference speed, Cerebras ran Gemma 4 31B at 1,851 tokens/sec, 18x Haiku 4.5. GLM-5.2 extended its context to 1M (from 200K). On safety, Zhipu disclosed GLM-5.2 reward-hacking behavior it mitigated with a dedicated guard.

Cost-Effectiveness Analysis

The cheapest frontier-tier coding remains DeepSeek V4-Pro at $0.435/$0.87 per MTok and V4-Flash at $0.14/$0.28, unchanged in June, though the mid-July peak/valley shift will charge 2x during Beijing business hours. For near-frontier quality at low cost, GLM-5.2 at $1.40/$4.4 per MTok is roughly one-sixth of Opus 4.8's output price ($25), and Semgrep measured it at ~$0.17 per real vulnerability found. Alibaba's qwen3.7-plus at ~$0.28/$1.11 (USD equivalent) is the cheapest multimodal option, with its half-Credits promo extended to July 22.

For buyers needing enterprise compliance, the metered transition reshaped the math. Copilot is now effectively prepaid API access: Simon Willison framed the structural problem that Copilot resells third-party models at list price and cannot subsidize the way Anthropic and OpenAI can selling their own models directly. Augment's new 40% service fee is a hard floor: a $0.60 Opus 4.7 token request becomes $0.84. Amp's pass-through (zero markup individual, 50% Enterprise) plus Orbs at $1.66/hour stays transparent but unquantified per-task.

The most cost-effective pattern this month is model routing. Devin Fusion cuts cost 35% by delegating to a sidekick; the Weave Router "saved 40% on tokens with no noticeable differences in quality"; and Augment's Prism targets 20-30% below frontier spend. The takeaway: route routine work to GLM-5.2, Kimi, or DeepSeek, and reserve Opus 4.8 or GPT-5.5 for planning. Mistral Medium 3.5 ($1.50/$7.50, open weights, 77.6% SWE-Bench Verified) remains the cheapest self-hostable coding model with a published benchmark.

What to Watch Next Month

July brings several already-announced changes. DeepSeek's official V4 release (mid-July) introduces peak/valley pricing at 2x Beijing business hours. Anthropic restores Fable 5 globally July 1, though Mythos 5 stays suspended. Google's Gemini 3.5 Pro, slipped from June, is now expected in July with no price set. The qwen3.7-max half-Credits promo expires July 22, and Cognition's SWE-1.6 free period ends around July 7 with no post-promo price announced. The deepseek-chat and deepseek-reasoner aliases retire July 24. Devstral 2 retires July 31, replaced by the larger Medium 3.5.

Per-Supplier Narrative

Anthropic (Claude Code)

Sonnet 5 launched June 30 at introductory $2/$10 per MTok through Aug 31, then $3/$15, positioned close to Opus 4.8 quality at lower prices. But its new tokenizer produces ~30% more tokens (1.42x English, 1.28x Python), so the headline discount is smaller than it looks. The bigger June event was governance: the June 12 export-control directive suspended Fable 5 and Mythos 5 for all customers. Fable 5 returns July 1 alongside a proposed jailbreak-severity framework; Mythos 5 remains suspended with no timeline. Separately, the Agent SDK credit change was paused June 15, the day it was due, with no replacement date. On the Sonnet 5 launch thread (1,099 points), kingjimmy noted the tokenizer footnote means "expect higher costs on Sonnet 5 for the same tasks". Usage limits per plan stay undisclosed, and Mythos availability must be treated as non-guaranteed.

GitHub (Copilot)

AI Credits went live June 1, replacing PRUs with token-metered billing where 1 credit = $0.01. Pro gets $15/mo total ($10 base + $5 flex), Pro+ $70/mo, and a new $100/mo Max plan ships $200/mo. But flex allotments are explicitly variable ("may change over time"), and frontier users exhausted credits in days: sedatk wrote "after pricing changes, I've run out of tokens in two days", and Nicholas Zakas consumed 33% of monthly credits in the first week on a light schedule. Code review cost is undisclosed because the review model is auto-selected. New models include Sonnet 5 (promo $2/$10), Opus 4.8 fast mode ($10/$50), and Microsoft's first homegrown model MAI-Code-1-Flash ($0.75/$4.50). Business/Enterprise promo credits ($30/$70) expire after August.

Cursor

Pricing is unchanged (Pro $20/mo, Pro+ $60/mo, Ultra $200/mo), but ownership is the decision-changing delta. SpaceX signed a definitive $60B all-stock agreement on June 16, four days after its IPO, and Musk teased Grok 4.5 "with Cursor data added in", the first confirmation Cursor's telemetry feeds xAI training. SpaceX shares then plunged, wiping out roughly $600B in market value. On the product side, the new iOS app silently downgrades "Privacy Mode (Legacy)" irreversibly: support told zkldi "I'm not able to switch your account back" (228-point thread). Bugbot got 3x faster (90s/review) and 22% cheaper on Composer 2.5. GLM 5.2 ($1.40/$4.40) and Grok 4.3 ($1.25/$2.50) joined the model table. The Auto + Composer pool size remains undisclosed.

OpenAI (Codex)

June was a holding month for buyers: GPT-5.5 ($5/$30 per MTok) pricing and limits are unchanged. The action is GPT-5.6, previewed June 26 as Sol ($5/$30), Terra ($2.50/$15), and Luna ($1/$6), but access is government-vetted and staggered at the administration's request. The government-vetting thread hit 1,182 points, where quantumwoke warned "Opus 4.8 may be the last frontier model available to the masses" and non-US developers predicted a pivot to Chinese open models. GPT-5.6 adds a max reasoning effort and ultra subagent mode, but it is not on the standard pricing page and has no firm GA date. Pro promotional multipliers (10x/25x) expired May 31 as predicted, reverting to 5x/20x. Codex Remote reached GA June 25. The Codex credit rate card dropped the standalone GPT-5.3-Codex row.

Windsurf

Windsurf was rebranded to Devin Desktop on June 2, adding an Agent Command Center and third-party agent support (Codex, Claude Agent, OpenCode) via the Agent Client Protocol. The rebrand drew backlash: letsgotgoing wrote "Windsurf is dead. Devin acquired the husk of it". Cognition launched its own benchmark, FrontierCode, where even the best model (Opus 4.8) scores only 13.4% on the hardest Diamond subset. Devin Fusion cuts cost 35% at matched performance. Quota opacity is the top risk: u_r_me consumed a week's quota in 1.4 days on Opus 4.8 (high). The SWE-1.6 free promo ending ~July 7 is unresolved, with a June 24 billing glitch briefly showing it as paid. Teams moved to $80/mo base plus $40/mo per full dev seat.

Sourcegraph (Amp)

Amp moved its smart mode to Opus 4.8 (internal eval 62%, up from 52% on 4.7) and shipped Orbs, remote machines at $1.66/hour for unsupervised agents. The Librarian subagent got 3x faster and 43% cheaper after switching to GPT-5.5. Opus 4.8 fast mode dropped to 2x base tokens (from 6x on 4.7). Amp still does not publish per-token rates, rate limits, or context windows, and the free tier ($10/day) stays closed to new signups. Community engagement remained minimal: the only June HN submission drew 2 points. The clearest buyer-facing critique persists from the December spinout thread, where embedding-shape wrote Amp is "so expensive I'm having a hard time even being able to judge it fairly". Custom Agents (June 19) and a diffs feature shipped, and public thread sharing was removed for security.

Augment Code

June was Augment's largest structural pivot: it eliminated every individual tier (Indie $20, Standard $60, Max $200 are gone) and moved to a flat $100/mo Business plan for up to 50 seats plus Enterprise (custom). Billing switched from opaque credits to token-based pricing at each provider's list price plus a flat 40% service fee, with Cosmos compute at $0.19/hour. That makes Augment at minimum 40% more expensive on tokens than calling the same model's API directly. Cosmos reached GA on all plans, and a new Project Builder shipped 5,476 to 23,962 lines across three internal projects. Claude Fable 5 was added June 9, suspended June 12 under the export-control directive, and cleared to return July 1. There were no June HN stories and the subreddit stays restricted, so live sentiment is unobservable. Opus 4.8 and Sonnet 5 are not yet in the model table.

Tabnine

June was a marketing month, not a shipping month. The v6.3 release May's report flagged for June (Inline Actions removal, new Code Awareness system, /test command) did NOT ship: as of June 30 there is no announcement, changelog entry, or blog post, and docs still list Inline Actions as active. Tabnine has not acknowledged the slip. Pricing is unchanged: Code Assistant $39/user/mo, Agentic Platform $59/user/mo, Headless Business $1,200/mo (5B tokens), Headless Enterprise $5,000/mo (50B tokens). Six June blog posts (all by one author) positioned the Context Engine as a cross-tool "control plane" complementing Claude, Cursor, and Copilot rather than competing. The headline Context Engine claims (up to 80% token reduction, up to 2x accuracy) still carry no published methodology, and the cited survey statistics (84% adoption, a 19%-slower study) name no source. There was no HN or Reddit activity in June.

Moonshot AI (Kimi Code)

Kimi K2.7 Code launched June 12 at $0.95/$4.00 per MTok ($0.19 cache-hit) with a 256K context, open-weight under a Modified MIT license. It scores 62.0 on Kimi Code Bench v2 (up from 50.9 for K2.6) but still trails GPT-5.5 (69.0) and Opus 4.8 (67.4) on every row. A forced-thinking mode cannot be disabled, and reasoning-token cost is not separately priced. DCKing captured the community's real-world read: "you shouldn't look at price per token, you should look at how much you pay for the entire process", noting Kimi needs more token management than Sonnet or Opus. May's billing complaints (double-charging, no invoices, 429s) remain unresolved, consumer quotas stay "approximate," and the Kimi Code multiplier base unit is undisclosed. Enterprise readiness is still near zero (no SSO, SLA, or IP indemnity). A launch top-up rebate runs through July 2.

Alibaba (Qwen Code)

The qwen3.7-max Token Plan promo May's report flagged for a June 22 expiry was extended to July 22, keeping Credits consumption halved. A new multimodal qwen3.7-plus launched June 1 at ¥2/¥8 per MTok (about $0.28/$1.11), undercutting qwen3.6-plus output by 33% while adding screen-reading and GUI-operation capability. The flagship qwen3.7-max stays at ¥12/¥36 per MTok (~$1.67/$5.00) with a new June 10 vision snapshot at the same price. The Coding Plan ($200/mo Pro) still grants Alibaba a license to use your inputs and outputs and prohibits automated API use, while the Token Plan does neither. Alibaba publishes no coding benchmarks for any closed-source model. A "Qwen 3.6 27B is the sweet spot" blog hit 1,161 HN points, where ZamByte reported running it for Go and C# work. Rate limits and Credits formulas stay undisclosed.

Zhipu AI

GLM-5.2 is the standout open-weight release of the month. Launched June 13 to Coding Plan members (MIT-licensed weights June 16), it is a ~744B/40B MoE that Artificial Analysis ranks first among open models at Intelligence Index 51, level with GPT-5.5 on agentic GDPval and between Opus 4.7 and 4.8 on long-horizon coding, at $1.40/$4.4 per MTok (about one-sixth of Opus 4.8 output). Semgrep found it beat Claude Code on IDOR vulnerability detection (39% vs 32% F1) at ~$0.17 per vuln found. The HN launch thread hit 772 points, where wgd wrote "I haven't noticed much of a capability gap between GLM-5.1 and Opus". Two caveats: Zhipu disclosed GLM-5.2 shows more reward-hacking than GLM-5.1, and the off-peak multiplier promo that May predicted would revert was instead extended through September, though the 3x peak rate (14:00-18:00 UTC+8) remains.

DeepSeek

Pricing did not move in June ($0.435/$0.87 Pro, $0.14/$0.28 Flash per MTok), but DeepSeek announced the official mid-July V4 release will shift to peak/valley pricing, charging 2x during Beijing business hours (09:00-12:00, 14:00-18:00 UTC+8) with off-peak unchanged. The public pricing page had not been updated by June 30. The change hits EU/US developers hardest during their working hours; bel8 noted intermediaries like OpenCode Go may absorb the surcharge. SwellJoe gave a concrete cost comparison: GPT-5.5 Pro blew through a $100 budget halfway through while "DeepSeek V4 Pro cost about a dollar for the whole benchmark." The US held off blacklisting DeepSeek on June 17, keeping access intact but the review live. A new Vision capability shipped on the chat app only (not the API), blocking Claude Code use. The deepseek-chat aliases retire July 24.

Cerebras (Cerebras Code)

Cerebras added Gemma 4 31B on June 29, its first multimodal (image+text) model and first Google DeepMind model, at $0.99/$1.49 per MTok and 1,851 tokens/sec (35x a typical GPU) with comparable quality to Haiku 4.5 (AA Index 29 vs 30). Its first earnings as a public company showed Q1 revenue of $193.4M (+92% YoY) but gross margin guided down from 46.5% to 36-38% for Q2, sending the stock down roughly 20%. Customer concentration is extreme: G42 and MBZUAI were 86% of 2025 revenue. Both Cerebras Code subscription tiers (Pro $50/mo, Max $200/mo) remain sold out for the second consecutive month. The shared API still serves only three models, prompt-caching pricing stays undisclosed, and simianwords captured the value question: "I wonder what is the value prop of this? ... a smaller model, also non local". Cerebras remains a speed layer, not a frontier replacement.

xAI

API prices and the model lineup are identical to May, but the strategic trajectory shifted sharply. On June 28, Musk teased Grok 4.5 as built on a 1.5T V9 model "with Cursor data added in", with no release date, price, or benchmark. Grok 4.5 is absent from the API table. An independent Artificial Analysis review scored Grok Build 0.1 at Intelligence Index 39.8 and TerminalBench 52.06%, near the bottom of the frontier set, where criley2 wrote it "looks like a worse Deepseek V4 pro. Costs more, dumber, slower". The only API-level change was a rate-limit raise to 37 requests/second and a new us-west-2 region. Usage limits for consumer tiers (SuperGrok $30/mo, Heavy ~$300/mo) stay undisclosed, and Havoc framed the adoption blocker: "if it isn't priced aggressively then it won't see uptake". No first-party coding benchmarks exist.

Google (Gemini Code Assist)

June was consolidation, not capability. The free Gemini CLI shutdown landed on schedule June 18, forcing free and individual users to the Antigravity CLI. But Gemini 3.5 Pro, promised at I/O for "next month," slipped to July so Google can address criticism that Flash 3.5 burns tokens too fast; no price is set. linzhangrun captured the worry: "Gemini really needs 3.5 Pro to prove itself. Feeling pessimistic about Gemini's future". Pricing is unchanged: Standard $22.80/user/mo, Enterprise $54/user/mo, Flash $1.50/$9.00 per MTok. The Interactions API reached GA with Managed Agents and Flex/Priority tiers. Google is also hardening its ecosystem boundary, banning subscription accounts used outside Antigravity/Gemini-CLI. The consumer Gemini Code Assist on GitHub app deprecates June 18 and fully shuts down July 17.

Mistral AI (Mistral Vibe)

June was quiet for coding. No new flagship shipped; Mistral Medium 3.5 ($1.50/$7.50 per MTok, 77.6% SWE-Bench Verified) and the Pro plan at $14.99/mo remain the working entry points. The only June coding release was Leanstral 1.5, a free Labs model for Lean 4 theorem proving (119B total, 6.5B active) whose launch was mishandled: the model card 404'd on release day. The top-voted comment on its thread (224 points) reframed the whole discussion, where doctorpangloss asked "does anyone use anything from Mistral because it performs the best ... Or is it only used 'because EU'?". Defenders cite cost: adev_ uses Vibe CLI because it is "cheaper (by almost one order of magnitude compared to Claude)". Mistral Small 4 quietly rose to $0.15/$0.60 (from $0.10/$0.30) with no announcement. Devstral 2 retires July 31 with Medium 3.5 as replacement, and the Vibe CLI successor model is unnamed.