The 2026 model lineup, decoded
OpenAI's naming has become its own usability problem. As of August 2026 there are nine models selectable inside ChatGPT, and the difference between picking the right one and the wrong one is larger than the difference between ChatGPT and its competitors. Here is the whole board:
| Model | Released | Built for | Who gets it |
|---|---|---|---|
| GPT-5.6 SolFlagship | 9 Jul 2026 Newest | Complex work across coding, research, science, cybersecurity, computer use and design | Paid plans; Sol Pro mode on Pro & Enterprise |
| GPT-5.6 TerraBalanced | 9 Jul 2026 | Everyday heavy lifting at roughly GPT-5.5 quality for less | Free & Go in Work/Codex; selectable on Plus+ |
| GPT-5.6 LunaFast & cheap | 9 Jul 2026 | Speed-first chat; the model behind Free's unlimited text | All tiers |
| GPT-5.5 Instant | 28 May 2026 | Natural conversational tone and pacing | All tiers |
| GPT-5.4 Thinking | 5 Mar 2026 | Extended reasoning and coding | All paid; Free via the Thinking toggle |
| GPT-5.4 mini | 18 Mar 2026 | Cost-efficient reasoning | Free & Go; fallback when paid limits hit |
| o3-pro | 10 Jun 2025 | Longest deliberation for maths and science | Pro and Team |
| o4-mini | 16 Apr 2025 | Fast, cheap reasoning on maths, code and images | All paid tiers |
| GPT-4.1 | 14 May 2025 | Coding with strict instruction-following | Plus, Pro, Team |
The one setting most people miss: GPT-5.6 exposes a reasoning effort dial, not just a model picker. Sol at low effort is not the model the benchmarks describe. If a task matters, raise the effort - and on Pro or Enterprise, use Sol Pro.
What GPT-5.6 actually improved
The July release was pitched on efficiency - "more intelligence from every token" - rather than a new capability class. The jumps that matter are in agentic and security work, not conversation:
GPT-5.6 Sol vs GPT-5.5, per OpenAI's own release figures. The ExploitBench jump - 47.9% to 73.5% - is the largest single-generation move in the set.
The other quiet addition is Programmatic Tool Calling, which lets the model orchestrate several tools in one pass instead of round-tripping through the API for each call. In practice that is why agent runs feel noticeably less sluggish than they did in the spring.
ChatGPT Work, and the death of Atlas
The biggest product change of 2026 landed the same day as GPT-5.6. ChatGPT Work (9 July 2026) is an agent that takes a goal, pulls from your connected apps, breaks the job into steps and returns finished artefacts - spreadsheets, decks, documents, working web apps - after running for as long as hours. It is built on Codex and GPT-5.6.
At the same time OpenAI began sunsetting Atlas, its standalone browser, folding the agentic browsing into ChatGPT's own built-in browser and a Chrome sidebar extension. That is the right call - a separate browser was always a hard ask - but it is also a reminder of how quickly OpenAI abandons surfaces. If you built a workflow on Atlas, you rebuilt it in July.
Watch the meter. ChatGPT Work is priced by usage, scaling with task complexity, in the same way Codex is - not covered by your flat subscription. A long agent run is not free just because you pay $20 a month.
Plans, prices and what you actually get
Six tiers now, and the headline price tells you very little - the real currency is message allowance and deep research quota.
| Plan | Price | What you get | Best for |
|---|---|---|---|
| Free | $0 | Unlimited text on GPT-5.6 Luna; roughly 10 top-model messages per 5 hours; ~5 deep research runs a month; limited uploads, images, voice, memory and Codex | Casual use |
| Go | $8ad-supported | ~10× Free's messages, full memory, expanded uploads and tools. No Thinking, deep research or custom GPTs | Light daily use on a budget |
| Plus | $20 | GPT-5.6 Sol at medium+ effort, ~25 deep research runs, projects, scheduled tasks, custom GPTs, ChatGPT Work and Codex agent | Most people Sweet spot |
| Pro | $100–200 | 5×–20× the usage, Sol Pro, ultra parallel-agent mode, unlimited Thinking, 125–250 deep research runs, maximum Codex tasks | Heavy agent and code workloads |
| Business | $20–25per seat, min 2 | No training on your data by default, admin console, SAML SSO, business connectors, unlimited standard messages | Small teams |
| Enterprise | Custom | Data residency in seven regions, SCIM and EKM, custom retention policies, 24/7 support with SLAs | Regulated orgs |
For developers, the API is where GPT-5.6's efficiency story shows up most clearly: Sol at $5 / $30 per million input/output tokens, Terra at $2.50 / $15, and Luna at $1 / $6. Luna in particular reprices a lot of high-volume work that was previously off the table.
Where ChatGPT is strongest
- Breadth. Nothing else does frontier reasoning, image generation, voice, video, code execution, browsing, scheduled tasks and a third-party GPT marketplace in one subscription. Rivals win categories; ChatGPT wins the average.
- Agentic follow-through. BrowseComp at 90.4% is the number to look at. Long multi-step web research is where GPT-5.6 has pulled clearly ahead, and ChatGPT Work converts that into files rather than transcripts.
- The ecosystem effect. ~900M weekly users as of February 2026, 1B monthly by June, and roughly 77% of consumer generative-AI traffic. Every integration gets built for ChatGPT first.
- Model granularity. Being able to choose Sol/Terra/Luna and an effort level is genuine cost control - if you understand it.
- Enterprise maturity. Seven-region data residency, SCIM, EKM and contractual SLAs are further along than most rivals.
Where it falls short
- The limits are deliberately murky. "Expanded", "extended", "5× more usage" - OpenAI rarely publishes hard numbers, and the caps move. You find out you have hit one mid-task.
- Metered agents on top of a subscription. Paying $200 a month and then paying again per agent run is a hard sell, however defensible the economics.
- Prose still reads like ChatGPT. The house voice - hedged, tidy, fond of triads - is recognisable, and Claude remains the stronger long-form writer out of the box.
- Nine models is too many. Legacy o3-pro, o4-mini and GPT-4.1 still sit in the picker alongside three GPT-5.6 variants. Most users will never pick correctly.
- Surfaces get retired fast. Atlas was launched and sunset inside a year. Build on the API and the core app, not the newest shiny thing.
- Context length isn't the lead. Gemini's 1M-token window is still the one to beat for whole-codebase or whole-corpus work.
- Ads arrive at $8. The Go tier is ad-supported - worth knowing before recommending it to a client.
ChatGPT vs the alternatives
The gap at the top has narrowed to the point where model choice matters less than fit. Here is the field as it stands in August 2026:
| Tool | Flagship model | Price | Where it wins | Rating |
|---|---|---|---|---|
| ChatGPTOpenAI | GPT-5.6 Sol9 Jul 2026 | Free / $8 / $20 / $100–200 | Breadth and agents. Best all-round capability, deepest integration ecosystem, strongest long-horizon web research (BrowseComp 90.4%). | 4.6/5 |
| ClaudeAnthropic | Claude Fable 5 · Opus 5Opus 5: 24 Jul 2026 | Free / $20 Pro / $100+ Max | Writing and code quality. The best long-form prose out of the box and a strong coding agent; Opus 5 lands near frontier performance at about half the price of Fable 5. | 4.5/5 |
| GeminiGoogle | Gemini 3.1 Pro19 Feb 2026 | Free / $4.99 Plus / $19.99 Pro / Ultra | Context and distribution. Up to 1M tokens, 77.1% on ARC-AGI-2, and it is already inside Gmail, Docs, Android and NotebookLM. | 4.5/5 |
| GrokxAI | Grok 4.3updated 1 Aug 2026 | Bundled with X Premium+ / SuperGrok | Price and real-time. $1.25 / $2.50 per million tokens undercuts every frontier rival, with live X data and a looser content policy. | 4.0/5 |
| Microsoft CopilotMicrosoft | Multi-modelincl. DeepSeek V4 trials | Bundled with Microsoft 365; Cowork metered | Where the work already is. Native in Word, Excel, Teams and Windows, with tenant-level governance IT already understands. | 4.0/5 |
| DeepSeekDeepSeek | DeepSeek V4 | Free app; near-commodity API | Cost and control. Open weights you can self-host, at a fraction of frontier pricing - good enough that Microsoft has trialled it inside Copilot. | 4.1/5 |
| PerplexityPerplexity AI | Multi-model + Comet | Free / $20 Pro / Max | Sourced answers. Citation-first search where you can check every claim - still the cleanest research surface, and the #2 consumer AI destination by traffic. | 4.2/5 |
| Le ChatMistral AI | Mistral open-weight family | Free / low-cost Pro | European data residency. Open weights, EU hosting and no US cloud dependency - a procurement answer more than a capability one. | 3.8/5 |
What the table actually tells you
ChatGPT and Claude have swapped roles at the margins. ChatGPT used to be the better writer and Claude the better reasoner; in 2026 it is closer to the reverse. If your output is prose a human will read end-to-end, try Claude first. If your output is a task completed across ten tools, ChatGPT.
Gemini's advantage is that you already have it. Google AI Plus at $4.99 is the cheapest credible frontier access anywhere, and for anyone living in Gmail and Docs the integration alone can outweigh a benchmark gap.
The price floor has collapsed. Grok at $1.25/$2.50 per million tokens and DeepSeek V4 with open weights mean "good enough" intelligence is now close to free. That pressure is exactly why GPT-5.6 Luna exists at $1/$6.
Which one should you actually pay for?
ChatGPT Plus - $20
Sol, deep research, projects, custom GPTs, agent and Codex. Still the best single-purchase answer.
Claude Pro
Cleaner default prose and less hedging. Opus 5 gets you near-frontier quality without the top-tier price.
Gemini 3.1 Pro
1M tokens of context, and it is already wired into Drive, Docs and Gmail.
Perplexity Pro
Citation-first by design, so you can verify every claim rather than trusting a summary.
Grok 4.3 or DeepSeek V4
An order of magnitude cheaper per token, and for classification or extraction the quality gap barely registers.
Copilot
Weaker as a chatbot, but it sits inside Excel and Teams with the governance IT already signed off on.
ChatGPT Pro
Ultra parallel agents and maximum Codex tasks. Budget for the metered usage on top.
Mistral Le Chat
Open weights and European hosting - often the only option that clears procurement.
Pros and cons
What we liked
- GPT-5.6 Sol is at or near the frontier on reasoning, coding and computer use
- BrowseComp 90.4% - long multi-step web research is a clear lead
- ChatGPT Work returns finished decks, sheets and apps, not transcripts
- Three price/performance tiers plus an effort dial gives real cost control
- Unmatched ecosystem: custom GPTs, connectors, Codex, voice, image, video
- Luna at $1/$6 per million tokens makes high-volume API work viable
- Serious enterprise controls - seven-region residency, SCIM, EKM, SLAs
What frustrated us
- Usage limits are vague by design and change without notice
- Agent and Codex usage is metered on top of the subscription
- Nine selectable models, including three legacy ones nobody should pick
- Default prose is still recognisably ChatGPT-shaped
- Atlas launched and was sunset within a year
- The $8 Go tier is ad-supported
- Beaten on context length by Gemini and on price by Grok and DeepSeek
FAQ
What is the newest ChatGPT model?
GPT-5.6, released 9 July 2026, in three tiers: Sol (flagship), Terra (balanced, roughly GPT-5.5 quality) and Luna (fastest and cheapest). Pro and Enterprise users also get a higher-quality Sol Pro mode.
Is ChatGPT Plus worth $20 a month?
For most people, yes - it is the tier where Sol, deep research, projects, scheduled tasks, custom GPTs and the agent all unlock. Step up to Pro only if you are running long agent jobs or heavy Codex workloads; the jump from $20 to $100–200 buys usage, not new capability.
Is the free version good enough?
For casual questions, yes. Free gives unlimited text chat on GPT-5.6 Luna, but roughly 10 top-model messages per five hours and about five deep research runs a month. If you regularly hit those walls, $8 Go or $20 Plus will pay for itself in saved friction.
ChatGPT or Claude?
ChatGPT for breadth, agents and integrations; Claude for long-form writing and code you will read carefully. Anthropic's Opus 5 (24 July 2026) reaches close to its frontier Fable 5 model at around half the cost, which makes Claude cheaper than it used to be at the quality tier most people need.
ChatGPT or Gemini?
Gemini 3.1 Pro wins on context (up to 1M tokens) and on being already present in Gmail, Docs and Android - and Google AI Plus at $4.99 is the cheapest frontier-adjacent access going. ChatGPT wins on agentic follow-through and the third-party ecosystem.
What happened to the Atlas browser?
OpenAI began sunsetting it in July 2026 and moved the agentic browsing into ChatGPT's own built-in browser plus a Chrome sidebar extension. Nothing was lost functionally, but it is a reminder that OpenAI retires surfaces quickly.
Does OpenAI train on my conversations?
On Business and Enterprise plans, no - that is the default. On consumer plans it depends on your data-controls setting, which you can turn off in settings. Enterprise adds regional data residency, SCIM provisioning and encryption key management.
Verdict
ChatGPT is still the default, and in 2026 it earned that rather than inherited it. GPT-5.6's gains are concentrated exactly where the market has moved - agents, browsing, computer use, security work - and ChatGPT Work is the clearest expression yet of the shift from "answer my question" to "finish my task". At 900 million weekly users and roughly three-quarters of consumer AI traffic, the ecosystem advantage compounds every quarter.
The reservations are about pricing honesty rather than capability. Opaque limits, metered agents stacked on a subscription, a model picker that has become a usability tax, and a habit of retiring surfaces faster than users can adapt. Meanwhile Claude has closed the reasoning gap and pulled ahead on prose, Gemini owns context length and distribution, and Grok and DeepSeek have dragged the floor price of competent intelligence close to zero.
Buy Plus. Keep a Claude or Gemini tab open for the jobs ChatGPT is second-best at. That combination costs $40 a month and beats any single subscription at any price.
Still the one to beat - but no longer the only one worth having
Frontier reasoning, the best agentic follow-through in the category, and an ecosystem nothing else approaches. Held back by opaque usage limits, metered agents on top of subscriptions, and a model lineup that badly needs pruning.