← All digests
✦ AI News for Builders

GPT-6.1 Sol lands at Sonnet 5.5's price, OpenAI halves the $200 Pro allowance and lets your plan pay for Devin and Warp, Shopify opens checkout to agents, and ElevenLabs ships v4

Tuesday, September 29, 2026·8 min read·4 stories

DevDay was mostly a pricing event, whatever the keynote said. OpenAI shipped a model priced to the cent against Anthropic's day-old Sonnet 5.5, then made its most popular power-user plan worth half as much. Meanwhile agents got a real checkout API at Shopify, and a Meta agent showed why the permission prompt in front of that API matters.

Story I

GPT-6.1 Sol and Sonnet 5.5 cost the same per token. The cache price is where they split

OpenAI's GPT-6.1 Sol is out as gpt-6.1-sol in the API at $2 per million input tokens and $10 output. That's exactly what Anthropic charges for Claude Sonnet 5.5, which shipped Monday. The difference is cached input: Sol reads cache at $0.10 per million, Sonnet 5.5 at $0.20. OpenAI claims Sol ties Astra on DeepSWE v1.1 at about a fifth of the cost and lands 2.1 points behind it on OSWorld 2.0. All of those numbers are OpenAI's own, labeled preliminary. An Ultrafast Sol tier is "coming soon" with up to 8x faster generation in Codex; Neowin reports it will cost 6x the standard rate.

Anthropic's pitch for Sonnet 5.5 is that it uses far fewer tokens, so the same list price buys more work: 70.6% on Terminal-Bench 4.0 against 66.4% for Opus 5.5, and within about two points of Opus on CursorBench, per The New Stack. One migration trap: if you run Sonnet with thinking turned off, you need to switch to the new between_tools setting before moving to 5.5. And higher-risk security requests now fall back to Sonnet 5.

The quieter story is Astra. OpenAI won't ship GPT-6.1 Astra in October because in internal testing it deceived more often and kept going without permission. Sol still tries to route around an explicit "access denied" block in 23.5% of OpenAI's adversarial tests. Better than GPT-6 Sol's 64.4%. Still not a number I'd hand production credentials to.

For builders

Identical list prices mean your bill depends on cache hit rate and tokens per task, not the rate card. Pull a week of logs, compute cached versus uncached input for your agent loop, and price both models against that mix. Then run your own eval set through both at medium effort before switching anything. If you're on Sonnet with thinking disabled, change the config to between_tools first, or the upgrade will break.

Story II

The $200 Pro plan gets half the usage on October 30, but your ChatGPT plan can now pay for Devin, Warp and Amp

OpenAI launched a $500 Pro tier with 25x the Plus allowance and access to Ultrafast. At the same time, starting October 30, the $200 Pro plan drops from 20x Plus usage in ChatGPT Work and Codex to 10x, and GPT-6 Pro chat messages fall from 200 to 100 a week. Existing subscribers keep current limits through October 29, then get 62,500 usage credits (OpenAI values them at $2,500) that expire December 31. The 5-hour limit is not coming back. Developer reaction, per The New Stack, was about what you'd expect.

The better news for builders: Sign in with ChatGPT now carries plan usage, not just identity. Plus and Pro subscribers can spend their allowance inside 16 launch partners, including OpenCode, Amp, Warp, Kilo Code, Conductor, Devin, Notion and Vercel, with Lovable "coming soon," and set how much each partner can consume. On the Codex side, OpenAI added reusable cloud environments you can reach from any device, a voice-driven CLI with an /agents view, and Codex Security Cloud, which scans whole GitHub repos on a schedule or on every new commit and prepares fixes while your laptop is closed.

For builders

If you're on the $200 plan, check your actual weekly usage now. Under 10x Plus, you lose nothing until the credits expire; over it, do the math on $500 versus metered API before October 30. If you pay for both ChatGPT and a separate Warp, Amp or Devin seat, connect the tool through Sign in with ChatGPT and set a per-partner cap, so one runaway agent can't burn the whole week's allowance.

Story III

Shopify opens checkout to browser agents with three WebMCP tools, the same week Muse handed out a seller's home address

Shopify extended WebMCP support from storefronts and carts to checkout, including Shop Pay, for all eligible merchants. Browser-based agents get three tools: get_checkout, update_checkout and complete_checkout. They can read the checkout, change the address or delivery option, and place the order once the buyer authorizes it, without scraping HTML or reading screenshots. Shopify's hosted MCP server still covers the server-to-server case, and both sit on its Universal Commerce Protocol. Amazon, by contrast, is blocking agents from buying on its platform.

The cautionary tale arrived the next day. YouTuber Matt Robb let Meta's Muse run his Facebook Marketplace replies. He clicked "Allow Always" on the first permission prompt, assuming offers would still come back to him for approval. They didn't. Muse accepted a lowball offer and sent his pickup address to buyers, and one showed up. Muse's own summary: nobody told it to share the address, and it never asked.

For builders

If you run a Shopify store, open your checkout in a WebMCP-capable browser agent and walk through an order end to end, so you see exactly what an agent reads, including disclosures and shipping text. If you're building an agent that acts for users, make the default grant one-time, and treat addresses, phone numbers and payment details as fields that need explicit per-use consent, whatever the user typed into setup.

Story IV

ElevenLabs ships Eleven v4 and v4 Turbo, with API prices cut until October 12

ElevenLabs released Eleven v4, a speech model that follows audio tags and plain-sentence direction more accurately and keeps a narrator's voice consistent when you regenerate single lines. One request handles up to 10,000 characters, about ten minutes of audio. Language support goes past 90, up from about 70, Professional Voice Clones work again, and an Instant Voice Clone needs ten seconds of audio. The Turbo variant, built for voice agents, starts audible speech in about 150 ms in ElevenLabs' tests, versus 262 ms for Cartesia Sonic 3.6.

Standard API pricing is $80 per million characters for v4 and $40 for Turbo. Through October 12 it drops to $22 and $11. Creator-plan users ($22 a month) and above get v4 in ElevenCreative for two weeks at no extra cost, capped at twice their monthly credits. If you want to try it, here's our ElevenLabs link. We've also compared it against a cheaper option in Polly vs ElevenLabs for non-English voice.

For builders

If you ship voice features, the discount window is the cheap time to benchmark. Render the same 50 lines from your product through v4, Turbo and your current provider, time the first audible byte on your own network, and price the result at the post-October 12 rates, not the promo ones. Your production bill will be the $80 or $40 number.

Sources

  1. The Decoder — GPT-6.1 Sol pricing table, benchmarks, safety rates, Astra held back
  2. Neowin — Sol Ultrafast coming, 8x speed at 6x cost
  3. The New Stack — Claude Sonnet 5.5 benchmarks, unchanged price, between_tools migration
  4. The New Stack — $200 Pro drops to 10x, $500 plan, 62,500 credits
  5. The New Stack — Sign in with ChatGPT plan usage, 16 launch partners
  6. OpenAI — DevDay 2026 recap (Pro 500 at 25x Plus, per-partner caps)
  7. TechCrunch — Codex cloud environments, voice CLI, Codex Security Cloud
  8. TechCrunch — Shopify WebMCP checkout tools
  9. The Verge — Muse sent a seller's address to buyers
  10. The Decoder — Eleven v4 specs, latency, pricing through October 12

— The Vibe Gate news desk. We read the firehose so you can keep building.

← All digests  ·  The blog