ChatGPT vs Claude vs Gemini in 2026: Capabilities, Pricing and Which to Use

The State of Play in September 2026

By September 2026, the major AI model families — OpenAI’s ChatGPT (GPT-5.6 Sol, Terra and Luna day to day, with GPT-6 Astra rolling out as the top tier from 3 September), Anthropic’s Claude (Fable 5.1, Opus 5, Sonnet 5, Haiku 4.5), Google’s Gemini (3.8 Flash and 3.1 Pro, with 3.5 Pro still unreleased after three missed dates) and, newly, Meta’s Muse Spark 1.3 — have converged to the point where raw benchmark scores alone are no longer a useful way to choose between them. The 2026 Stanford AI Index made that point in the spring, and the summer’s releases have only tightened the cluster.

What does matter is understanding where each model family has a genuine edge. The differences are real, but they show up in specific task types rather than overall intelligence. This article breaks down those differences so you can make an informed choice for your actual workflow.

For a deeper foundation in how these models work under the hood, see the AI Fundamentals course. For practical workflows using all of them, see Mastering AI Tools.

Capabilities Compared

Context Window

This used to be a differentiator. It isn’t any more. All three GPT-5.6 tiers carry a 1 million-token window (GPT-6 Astra nudges that to 1.05 million), as do Claude Fable 5.1, Gemini 3.8 Flash, Gemini 3.1 Pro and Muse Spark 1.3. Only the smallest models — Claude Haiku 4.5 at 200,000 tokens, for instance — sit below that. In practical terms, every frontier model can process an entire codebase, a lengthy report or a multi-document research set in a single interaction. Where they differ is how well they use the far end of that window: Meta claims 98.5% on the MRCR long-context test at 256–512K tokens for Spark 1.3, which is the strongest published figure I’ve seen.

Multimodality

Gemini remains the most fully multimodal: it takes text, images, audio and video in, and its media stack — Nano Banana 2 and Nano Banana Pro for images, Gemini Omni Flash for video — is built into the same product. ChatGPT supports text and image input and generates images through ChatGPT Images 2.0 (the gpt-image-2 model), which reasons about composition before it draws; it does not natively output audio or video. Muse Spark 1.3 accepts text, image and video input, and Meta shipped its first image model, Muse Image, in July. Claude supports text and image input but still lacks native image generation or speech.

Reasoning

All four now expose reasoning as a dial rather than a separate product. In ChatGPT the old Instant and Thinking modes have been folded into one model with a reasoning-effort slider. Claude Fable 5.1 adds per-message effort control in the API and is noticeably stronger than Fable 5 at high effort. GPT-6 Astra exposes five effort levels in the API, from low to max, and OpenAI ran its launch benchmarks at max — worth remembering when you read them. Gemini 3.8 Flash is pitched as the “most intelligent workhorse”: it takes smaller reasoning steps, calls tools iteratively and verifies its own work, trading more tokens for better accuracy, with thinking-level controls to rein it in. Muse Spark 1.3 ships in “max” and “xhigh” variants. The pattern from the spring holds: each wins on different kinds of reasoning, and there is no universal leader.

Code Generation

This is the most contested category and the one where the September numbers are freshest. On DeepSWE 1.1, Muse Spark 1.3 scores 75.4, Claude Opus 5 74.0 and GPT-5.6 Sol 73.0 — and OpenAI’s launch-day figure for GPT-6 Astra is 74.1, which lands squarely in the same pack (its chart left Muse off and used a 67.4 result for Fable 5.1 that the public leaderboard doesn’t agree with). On Terminal-Bench 2.1, Spark and Sol tie at 88.8, with Opus 5 at 86.7. Those gaps are small enough to be noise for most real work; there is no clear coding leader. Beyond benchmarks, Claude’s reputation for root-cause debugging rather than band-aid fixes is the thing developers keep telling me about, and Fable 5.1’s biggest gains are in exactly that long-running agentic coding. ChatGPT has the broadest language coverage and the most mature tooling, with Codex now running on GPT-5.6 and passing four million weekly developers by late May. Gemini provides Code Assist and Antigravity 2.0 for the Google Cloud crowd.

Agentic Capabilities

Every provider now has a first-party coding agent: OpenAI’s Codex, Anthropic’s Claude Code (on Fable 5.1 since 1 September, with roughly 60% fewer cybersecurity false positives), and Google’s Antigravity 2.0 (desktop app, CLI, SDK, subagents). Google also maintains the Agent Development Kit for building your own. Fable 5.1 adds readable progress updates between tool calls, which sounds minor until you’ve watched an agent run for forty minutes in silence. All four support function calling and tool use through their APIs. For a deep dive into building agent workflows, see the AI Agents course.

Speed and Latency

Each provider offers a spectrum of model sizes for different speed and cost trade-offs. GPT-5.6 Luna is OpenAI’s fast tier and has been the default for free and Go users since 6 August. Claude Haiku 4.5 is designed for speed, running about 40% faster than Sonnet 5. Google’s Flash line is the low-latency option, though be aware that 3.8 Flash deliberately spends more tokens on verification than its predecessors. For most production use cases, the lighter models from any provider are fast enough for real-time interaction.

Pricing

Pricing falls into two categories: consumer subscriptions and API usage.

Consumer Subscriptions

ChatGPT Plus is $20/month and defaults to GPT-5.6 Sol, with GPT-6 Astra rolling out to Plus, Pro, Business and Enterprise this week (Astra Pro on Pro and above); free and Go users get Luna. Claude Pro is $20/month, with Opus 5 and Sonnet 5 included and Fable 5.1 available through usage credits — it is only fully included on Max, Team Premium and premium Enterprise seats. Google AI Pro is $20/month and already has Gemini 3.8 Flash live in the app, AI Mode and Sheets. Meta is rolling Muse Spark 1.3 into Instagram, Facebook and Meta AI rather than selling a standalone subscription. At the consumer level, pricing is effectively identical — the choice is about features and which frontier model you actually get for the money.

API Pricing

A few important details. Fable 5.1 and GPT-6 Astra now share the top of the price list at $10/$50 — Astra is 2.5× Sol’s promo rate, so it is a deliberate choice rather than a default. On Fable 5.1, the cheaper cache reads change the arithmetic for anything agentic, where the same context gets re-read dozens of times. GPT-5.6 Sol’s $4/$20 is a promotion with an end date; budget on $5/$30 if your contract runs past November. Gemini 3.8 Flash’s introductory price doubles in January. And Meta’s contributor endpoint is the cheapest frontier compute on the market by a wide margin — but read the training clause twice before you put client data through it. For high-volume production workloads, Luna, Haiku 4.5 or 3.8 Flash are the sensible defaults.

Benchmark Performance

Independent evaluations confirm all four are frontier-class, usually within a few points of each other. Here is where each currently leads on the September numbers:

• Agentic coding (DeepSWE 1.1): Muse Spark 1.3 — 75.4 (vs Claude Opus 5 at 74.0, GPT-5.6 Sol at 73.0)
• Terminal tasks (Terminal-Bench 2.1): Muse Spark 1.3 and GPT-5.6 Sol tied — 88.8 (vs Opus 5 at 86.7)
• Long-context recall (MRCR, 256–512K): Muse Spark 1.3 — 98.5%
• Overall (Artificial Analysis Intelligence Index): Muse Spark 1.3 sits sixth of 636 models; among open-weight models, GLM-5.2 scores 51 to DeepSeek V4 Pro’s 44
• Video generation (August text-to-video leaderboards): Gemini Omni Flash, alongside Wan 3.0
• Agentic coding, OpenAI’s figure (DeepSWE 1.1): GPT-6 Astra — 74.1 at max effort. Same dead heat: Muse 75.4, Opus 5 74.0, Sol 73.0, and the public leaderboard has Gemini 3.8 Flash and Opus 5 at 74
• Computer use (OSWorld V2-Offline): GPT-6 Astra — 72.6% vs Sol 65.7%. Anthropic reports 77.9% for Fable 5.1 on a different OSWorld release, so the two aren’t comparable
• Science and maths (Astra, OpenAI-reported): GPQA Diamond 96.0% vs Fable 5.1 93.7%; Terminal-Bench Science 64.6% vs Fable 5.1 52.6%; FrontierMath Tier 4 97.6% vs Fable 5.1 87.8% — though Epoch AI notes OpenAI funded FrontierMath and has exclusive access to part of it
• ARC-AGI-3: GPT-6 Astra — 98.6% vs Opus 5 30.2%, run in a harness that keeps reasoning between turns, which OpenAI itself has shown can triple ARC scores without changing the model
• Frontier maths and CS: Astra solved ten open problems in August with Lean-4-verified proofs — still the most impressive single result of the year, and now something you can actually buy

Three cautions. First, every GPT-6 Astra number above is OpenAI’s own, run at max effort, on a chart that left Muse Spark off entirely; wait for independent runs before you move budget. Second, Anthropic’s own framing for Fable 5.1 is that it’s similar-or-better than Fable 5 at low and medium effort and much stronger at high effort, which is the sort of claim that only shows up in benchmarks if the benchmark was run at high effort. Third, benchmark leadership rotates with every release. The overall gap has narrowed so much that integration, features and workflow fit matter more than raw accuracy for most practical decisions.

Strengths and Weaknesses

ChatGPT (GPT-6 Astra / GPT-5.6 Sol / Terra / Luna)

Strengths: Broadest general capability across the widest range of tasks. Strong complex reasoning and multi-step problem solving. The most mature developer ecosystem, with Codex, extensive API tooling and community libraries. Reliable function calling and tool use. A clean line-up — three GPT-5.6 tiers plus GPT-6 Astra on top — that makes cost routing straightforward. Astra is state of the art on computer use, browsing and science by OpenAI’s own numbers. ChatGPT Images 2.0 is the best image tool for anything with text in it.

Weaknesses: Sol’s current price is promotional, so plan for the list rate. Can be conservative with refusals on edge-case queries. Data privacy requires an enterprise plan. GPT-6 Astra is 2.5× Sol’s price, its cyber capabilities are gated behind the Daybreak trusted-access programme after it crossed OpenAI’s “Critical” threshold, and OpenAI says its misalignment monitors may slow, pause or stop legitimate long-running tasks at launch — so build a checkpoint into anything you run on it unattended.

Claude (Fable 5.1 / Opus 5 / Sonnet 5 / Haiku 4.5)

Strengths: Best-in-class instruction following and long-running agentic coding, with a well-earned reputation for fixing the actual bug rather than the symptom. Strong on multistep research and document, spreadsheet and slide work. Strong emphasis on safety and alignment. Excellent writing quality with natural tone. Does not train on user data by default. Available on AWS and GCP. Transparent about uncertainty — more likely to say it doesn’t know than to confabulate.

Weaknesses: Fable 5.1 is the most expensive frontier model on list price and only fully included on the top subscription tiers. Still no native image generation or speech. Enterprise pricing (seat plus usage) can be complex. Some API sharp edges in 5.1: forced tool use now errors, and editing earlier turns invalidates thinking blocks. The June export-control suspension of Fable 5 — and this week’s music-publisher lawsuit naming Anthropic’s founders personally — are reminders that vendor risk is real for every provider.

Gemini (3.8 Flash / 3.1 Pro)

Strengths: Native multimodality across text, image, audio and video, with the strongest media stack (Nano Banana 2, Nano Banana Pro, Omni Flash). Deep integration with Google Workspace. Real-time web grounding reduces hallucination for research tasks. The cheapest frontier-class workhorse at the current 3.8 Flash intro price. Antigravity 2.0 and the ADK for agents.

Weaknesses: The Pro tier is stuck: 3.1 Pro is from February, and 3.5 Pro has missed three promised dates. Three Flash releases in three months is impressive velocity but hard to build against. 3.8 Flash’s verification habit costs tokens. Instruction following is still less precise than Claude’s, and polished professional writing generally trails both GPT and Claude.

Meta Muse Spark 1.3 — the fourth contender

Meta Superintelligence Labs’ closed frontier model is now impossible to leave off this list. Spark 1.3 arrived on 2 September, about four weeks after 1.2, with a 1M context, text, image and video input, and benchmark results that lead or tie on DeepSWE 1.1 and Terminal-Bench 2.1. It is also cheap — $1.25/$4.25 per million tokens on the standard endpoint, and far cheaper still if you accept the contributor terms. Strengths: price-to-performance, long-context recall, a monthly release cadence, and an open-weight sibling line (Muse Spark 1.2’s weights are open, and the ~30B-parameter Muse Glimmer runs on a laptop under Apache 2.0). Weaknesses: it is a month old, the enterprise story is thin next to the other three, the consumer surface is Instagram and Facebook rather than a serious work app, and the contributor endpoint’s training clause rules it out for anything confidential. Pilot it for cost-sensitive API workloads; don’t move your enterprise onto it yet.

Enterprise Features

All three providers now offer enterprise-grade security, and the differences are narrowing. Here is a summary of what matters for organisational buyers:

The practical guidance for enterprise buyers: if your organisation runs on Google Workspace, Gemini's native integration is a genuine advantage. If you use AWS or have specific compliance requirements around safety and alignment, Claude's enterprise plan and Constitutional AI approach may be the better fit. If you need the broadest developer ecosystem and most mature API tooling, ChatGPT Enterprise is the safe choice.

For organisations evaluating AI adoption at scale, the Corporate Training programme covers model selection, governance and implementation planning.

Safety and Alignment

All four take safety seriously, but with different approaches, and the summer produced an interesting convergence: every major lab now gates its most dangerous capabilities behind a trusted-access tier. Anthropic ships Fable 5.1 and Mythos 5.1 as the same model at two safeguard levels — Fable generally available, Mythos only via trusted-access programmes for cybersecurity and life-sciences work. Google’s Gemini 3.8 Flash Cyber is available only through “Fairwind”, its limited-access programme for governments and trusted partners. OpenAI classifies frontier models under its Preparedness Framework; GPT-6 Astra, released 3 September, is the first model over the framework’s “Critical” cyber threshold, so the public version refuses exploit discovery and advanced cyber work runs only through the Daybreak programme. OpenAI also says, in its own words, that Astra’s chain of thought is harder to monitor than Sol’s — an honest admission, and not a comforting one. Anthropic’s Constitutional AI approach, where models are guided by explicit principles rather than just human labels, remains its distinctive contribution.

In all cases, enterprise customers get encryption, SSO, audit logs and compliance guarantees. None of the providers use customer content to train their models by default — with the pointed exception of Meta’s opt-in contributor endpoint, which is cheap precisely because you’re paying in data. For a deeper understanding of how AI safety frameworks work, see the AI Safety glossary entry.

Which Model to Use for What

Here is a practical selection guide based on task type:

• Daily writing and email: Any lighter model — GPT-5.6 Terra or Luna, Claude Sonnet 5, or Gemini 3.8 Flash. Fast, cheap, good enough for routine work.
• Complex analysis and reasoning: GPT-6 Astra or Claude Fable 5.1 at high effort (GPT-5.6 Sol if the budget matters). GPT has the edge on breadth, Claude on precision.
• Coding and software development: Claude Fable 5.1 or Opus 5 for careful, root-cause work in Claude Code. GPT-5.6 Sol in Codex for breadth across languages and frameworks. Muse Spark 1.3 if cost per token is the constraint.
• Research with current information: Gemini 3.1 Pro or 3.8 Flash with grounding. Its real-time web search integration is still the most natural.
• Google Workspace workflows: Gemini 3.8 Flash, which is already live in Sheets. Native integration with Docs, Gmail and Drive is a genuine advantage.
• Long-form content and creative work: Claude Opus 5 or Fable 5.1 for voice and nuance. GPT-5.6 Sol for breadth and boldness.
• Agent and automation workflows: Claude Fable 5.1 for long-running, multistep work and document, spreadsheet and slide tasks. GPT-5.6 Sol for tool-use breadth. Gemini for Google-ecosystem agents.
• Image and video work: Gemini for Nano Banana 2 / Pro and Omni Flash video. ChatGPT Images 2.0 for anything with text or layout in it. See the AI for Creatives course for detailed workflows.
• High-volume production (APIs): GPT-5.6 Luna, Claude Haiku 4.5, Gemini 3.8 Flash at the intro price, or Muse Spark 1.3 — whichever your stack already supports.

The Mastering AI Tools course teaches you how to evaluate and select models for specific professional workflows, including hands-on exercises comparing outputs across all the major providers.

Frequently Asked Questions

Which AI model is best overall in 2026?

There is no single best model. ChatGPT (GPT-6 Astra on top, GPT-5.6 Sol as the everyday default) leads in breadth, Claude (Fable 5.1 / Opus 5) leads in coding and agentic work, Gemini (3.8 Flash / 3.1 Pro) leads in multimodality and Google integration, and Meta’s Muse Spark 1.3 leads on price-to-performance. The right choice depends on your specific tasks. Many teams use two or more for different parts of their workflow.

How much does ChatGPT vs Claude vs Gemini cost?

Consumer subscriptions are roughly the same at $20/month each, though which frontier model you get differs: ChatGPT Plus includes GPT-5.6 Sol with GPT-6 Astra rolling out this week, while Claude Pro puts Fable 5.1 behind usage credits. API pricing varies by tier. Lightweight models like GPT-5.6 Luna ($0.20/M input, $1.20/M output) and Gemini 3.8 Flash ($0.75/$3.75 until the end of 2026) are very affordable. Frontier models like Claude Fable 5.1 and GPT-6 Astra (both $10/M input, $50/M output) and GPT-5.6 Sol ($4/$20 on promotion) are best reserved for tasks that need their full capability.

Which model is best for coding?

It’s close. On DeepSWE 1.1, Meta’s Muse Spark 1.3 scores 75.4, GPT-6 Astra 74.1 (OpenAI’s own figure, at max effort), Claude Opus 5 74.0 and GPT-5.6 Sol 73.0 — a dead heat once you allow for uncertainty ranges; on Terminal-Bench 2.1, Spark and Sol tie at 88.8. Claude’s edge is in long-running agentic work and root-cause debugging, and Fable 5.1 pushed that further. GPT-5.6 handles a wider range of languages and frameworks through Codex. For most developers, any of them works well — test on your specific codebase.

Can I use multiple AI models together?

Yes, and many teams do. A common pattern is to use a fast, cheap model (GPT-5.6 Luna, Gemini 3.8 Flash or Claude Haiku 4.5) for drafting and classification, then route complex work to a stronger model (Fable 5.1, GPT-6 Astra or Gemini 3.1 Pro) for final output. This reduces cost while maintaining quality where it matters.

Which model is safest for enterprise use?

All the major providers offer enterprise-grade security with SOC 2, ISO, HIPAA readiness, SSO, encryption and data residency. None use customer data for training by default (Meta’s cheaper contributor endpoint is the explicit exception). The best choice depends on your existing cloud provider and compliance requirements.

How often do AI model comparisons change?

Constantly. Between May and September 2026 every provider shipped at least one new frontier model, and Google shipped three. This article reflects 3 September 2026 and is updated regularly. The selection principles — match model strengths to task type — remain stable even as rankings shift.