The AI Model Landscape
Seven platforms, the capabilities that are now table stakes, and how insurance leaders should choose and govern AI in a regulated environment. The capability gap between frontier models has narrowed. Workflow fit, ecosystem, governance, and strategic alignment should drive your choice, not fear of missing a feature.
Time to read: about 13 minutes
Last verified August 2026. Model versions and pricing change monthly, so confirm current details before acting.Universal capabilities
What is now table stakes
These capabilities are standard across all frontier models. They should not drive your platform choice, because every serious contender delivers them competently. For most insurance use cases, any frontier model will produce competent results. The differentiation lies in secondary characteristics: hallucination rate, ecosystem integration, citation practices, governance, and data sovereignty.
Multi-step reasoning
Every platform offers a thinking or deep-reasoning mode, and all score at the top of graduate-level knowledge tests.
Document analysis
Processing PDFs, spreadsheets, contracts, and images works well across every major platform.
Large context windows
Context windows of one million tokens, roughly 750,000 words, are available across the frontier tiers of Claude, Gemini, ChatGPT, and now DeepSeek.
Real-time web search
Integrated across all major platforms. No longer a differentiator on its own.
Persistent memory
Every platform offers some form of conversation memory that carries context across sessions.
Multimodal input
Vision and, on most platforms, audio and image generation are increasingly standard rather than distinguishing.
Seven platforms
The field at a glance
Each profile covers the company, the current models, the platform, and where it leads. Figures are current as of the verification date above. This edition adds Microsoft Copilot, because for many insurers the real question is not only which frontier lab leads, but which tool is already embedded in the environment your people work in every day.
The safety-first enterprise choice for serious analytical work
Models
The Opus line moved quickly through 2026: Opus 4.6 in February, 4.7 in April, 4.8 in May, and Claude Opus 5 in July, which closed most of the gap to the frontier tier while holding the prior price of roughly $5 per million input and $25 per million output tokens. Opus 5 anchors the Max plan. Sonnet 5 became the default for Free and Pro users and makes near-flagship agentic work a daily-driver expense. Opus 4.1 was retired on August 5, 2026, so any workflow still naming it should be migrated. Claude consistently leads on writing quality, clean production-ready code, low hallucination rate, and financial reasoning, which is why it tends to be the right tool for open-ended, multi-step analytical work rather than single-shot assistance.
New-user note. No native image generation. Occasional over-refusal of legitimate requests. The interface rewards deliberate setup and iteration rather than one-shot prompting.
A note on the Mythos tier and export controls
In June 2026 Anthropic released Claude Fable 5, the first public model from its Mythos-class tier, a step above the Opus line. Days later, a US export-control directive forced Anthropic to suspend access to Fable 5 and its sibling Mythos 5 over national-security concerns, and because it could not filter access in real time, both models went dark globally for about three weeks while other Claude models kept running. Access was restored on July 1, 2026, and a restricted Mythos configuration was later cleared for a small set of approved security and infrastructure customers. For insurance decision-makers, this is the clearest real-world case for multi-model redundancy: frontier capability is now entangled with export-control and national-security policy in ways that can interrupt access to a single model with little notice, even for paying enterprise customers.
Leads on writing quality, code generation, low hallucination, financial reasoning, enterprise trust
The broadest ecosystem and the largest user base
Models
The GPT-5 family ships near-monthly. The current generation is GPT-5.6, with named variants including Sol, Terra, and Luna tuned for different planning, building, and reviewing workloads. Pricing spans a wide range from the flagship Sol tier down to the low-cost Luna tier, and OpenAI cut those rates more than once through mid-2026, so the numbers move faster here than on any other platform and should be checked against current rates rather than assumed. Configurable reasoning depth, native computer use, and a broad multimodal span across vision, voice, image, and video generation remain strengths. ChatGPT leads on breadth of occupational knowledge and on the sheer range of integrations available around it.
New-user note. The model you are talking to can change mid-conversation, because the platform auto-routes between faster and deeper modes, and free-tier sessions may be quietly downgraded under load. Rapid deprecation of older versions can disrupt established workflows.
Leads on broadest ecosystem, consumer adoption, image and video generation, third-party integrations
The integration giant, strongest inside Google Workspace
Models
Google has iterated hard on its Flash workhorse tier, with Gemini 3.6 Flash and then 3.7 Flash reaching general availability through mid-2026, tuned for coding and agentic workflows at low cost, alongside a Flash-Lite option for high-volume automation. Notably, Google had not yet shipped a Gemini 3.5 Pro update to its top reasoning tier as of late August, and is generally seen as trailing Anthropic and OpenAI at the frontier while competing aggressively on price and Workspace integration. Gemini's native multimodal design and NotebookLM source-grounded research remain genuine strengths.
New-user note. The Auto model setting can silently route you from the Pro tier down to Flash. Quality can vary, and very long contexts can drift.
Leads on native multimodal, Workspace integration, price-performance, source-grounded research
The assistant already embedded where your people work
What it is
Microsoft 365 Copilot is an assistant layer built over frontier models and wired directly into the Microsoft 365 data graph: email, documents, meetings, and chat. Its defining strength is grounding, surfacing answers from your own organizational content while honoring existing permissions and sensitivity labels through Microsoft Purview. For organizations already standardized on Microsoft 365, it puts AI where work already happens, with governance that inherits the controls the organization already runs.
Where the seam is
Because Copilot is designed as a grounded assistant over your data, it is strong at retrieval and single-shot help and less suited to open-ended, dynamic, multi-step reasoning that has to hold and revise a long analytical chain. Teams doing serious iterative analytical work often find they reach the edge of that design and need a frontier model driven conversationally instead. The practical pattern many insurers land on is Copilot for broad grounded assistance across the Microsoft estate, and a frontier model such as Claude for the serious analytical work, rather than treating either as the single answer.
New-user note. The consumer Copilot at copilot.microsoft.com runs under a consumer privacy policy and is not appropriate for regulated or PII-bearing work. Only the licensed Microsoft 365 Copilot carries the enterprise data-protection commitments described in the governance section below. Confusing the two is a real exposure.
Leads on Microsoft 365 integration, grounded retrieval over your own data, inherited enterprise governance
The answer engine that cites its sources
Models and approach
Perplexity is not a single model but a retrieval pipeline that routes across frontier models and its own Sonar models, with a mode that runs a query across several models and shows where they agree and disagree. That cross-model design is itself a working illustration of the multi-platform thesis this guide argues. The company raised a Series E-6 round in early 2026 at roughly a $23 billion valuation, with revenue estimated in the range of several hundred million dollars annualized and growing quickly. It excels at real-time, cited fact retrieval, with focus modes for finance, academic, and other source types.
New-user note. It is a research tool, not a writing partner. Notably weaker for creative writing, long-form drafting, and complex code generation than Claude or ChatGPT.
Leads on cited research, factual accuracy, multi-model comparison, real-time data
The open-weight era gives way to a closed pivot
The pivot
The closed-model shift this landscape previously flagged as a rumor has happened. In April 2026 Meta launched Muse Spark under Meta Superintelligence Labs as a closed, API-only successor to Llama, its first frontier model not released as open weights since 2023, with a Muse family expanding since. Llama 4, Scout and Maverick, remains available and widely deployed, and the open-weights strategic lessons still hold as case studies: open models create pricing pressure on proprietary vendors, provide negotiating leverage, and, when self-hosted, keep sensitive data entirely within the security perimeter. But the forward-looking assumption that Meta would keep competing on the open-weights axis no longer holds cleanly, which weakens the vendor-leverage argument that assumed a strong open alternative would always exist.
New-user note. The Llama models are weights you run on your own infrastructure or access through third-party hosts, not a website you visit. Muse Spark is closed and API-only. Self-hosting either open Llama models or alternatives requires real GPU infrastructure and in-house capability.
Leads on data sovereignty (via open Llama), cost efficiency, customizability, vendor diversification
The efficiency disruptor that rewrote cost assumptions
Models
DeepSeek V4, a mixture-of-experts family under the MIT license, superseded the V3.2 generation. V4-Pro runs 1.6 trillion total parameters with roughly 49 billion active per token and a one-million-token context window, at output pricing around $0.87 per million tokens, keeping a substantial cost advantage over Western frontier models. The strategic signal is the point: frontier-class capability built for a fraction of the assumed cost, which is a genuine question for any technology roadmap that assumed frontier AI would stay expensive.
Critical caveat for regulated industry. Data sent to the hosted web app and API flows to servers in China, and multiple governments and agencies have restricted the app. The model also applies content filtering aligned to Chinese state positions. For an insurer, the only viable path is a self-hosted open-weights deployment under compliance review, never the hosted service with any policyholder or claims data.
Leads on cost disruption, efficiency innovation, open weights (MIT), strategic signal
Governance and enterprise plans
Choosing and governing AI in a regulated environment
For an insurer, the platform decision is only half the question. The other half is which plan tier you buy and how the data is governed, and that half is where the regulated-industry risk actually lives. This section is the part general capability comparisons leave out.
First question: is this decision even yours to make?
Before comparing plan tiers or vendors, establish what agreements already exist above you. A parent company, a group entity, or an existing enterprise contract with a model provider or with Microsoft may already cover affiliated and subsidiary organizations. Buying a standalone plan when a parent enterprise agreement already reaches your entity is redundant cost, split governance, and sometimes a procurement problem. The first move in any enterprise AI plan decision is to confirm whether you are actually making a fresh buying decision or operating inside an agreement that already exists. Only once that is settled do the feature, cost, and governance comparisons below matter.
The plan tier is a data boundary, not just a feature set
Across every major platform, the free or consumer tier and the licensed enterprise tier are governed by different contractual terms. The consumer tier typically runs under a consumer privacy policy and is not appropriate for regulated or PII-bearing work. The licensed enterprise tier carries the data-processing terms, tenant isolation, no-training commitments, and residency controls that regulated work requires. The single most common and most dangerous mistake is using a consumer-grade tier for work that involves claims data, policyholder PII, or other regulated content. The tier you choose is a governance boundary before it is a feature choice.
Protection and permission, with different controls for different work
Governance in a regulated shop is both protection and permission: it protects sensitive data and it grants people the room to work. That usually means different controls for different kinds of work: broad latitude for low-sensitivity personal productivity, tighter controls for finance and internal analysis, and the strictest handling for regulated, consumer-facing work involving PII. A particular concern in insurance is that conversational records and claims information can be exposed or interpreted out of context, which is exactly what tenant isolation, retention controls, and sensitivity labeling exist to contain. Microsoft 365 Copilot inherits these through Purview and enterprise data protection; the frontier labs provide their own equivalents at the enterprise tier.
Cost is a moving target that needs its own monitoring
Pricing across these platforms is volatile and genuinely hard to track. Rates change with little notice, vary by tier and by usage pattern, and shift often enough that even tools built to follow them fall behind. An organization running AI at any scale needs a recurring mechanism to monitor and understand what it is spending, the same way it tracks any other operating cost, rather than treating a price checked once as settled. The cost of operating a model is not a number you look up at purchase, it is a variable you manage over time.
A worked example: choosing a Claude plan tier
Claude's own structure illustrates the general principle. The models are identical across tiers, so you do not move up for smarter output, you move up when compliance, identity, or scale requires it. Claude Team is the self-serve tier at a published per-seat price, includes single sign-on, central billing, and a no-training commitment, and suits a department or pilot group that wants governed, shared access without a procurement cycle. Claude Enterprise adds the controls regulated organizations cannot skip: SCIM identity provisioning, audit logs, custom data-retention controls, customer-managed encryption keys, US-only inference, and HIPAA-ready configuration, with usage billed separately from the seat fee. The decision rule is one sentence: choose Team for governed shared access, and graduate to Enterprise when a compliance, identity, or scale requirement forces it, not before. Seat minimums and exact pricing shift, so confirm current terms on the provider's pricing page as part of the decision.
The strategic case
Why multi-platform fluency matters
The argument for multi-platform competency applies with particular force to insurance, and the events of the past year strengthened rather than weakened it.
Different models lead on different work
Claude leads on writing, clean code, and financial reasoning. Gemini competes on multimodal and price. ChatGPT leads on ecosystem breadth. Perplexity leads on cited retrieval. Copilot leads inside the Microsoft estate. No single platform dominates every dimension.
Access can be interrupted
The brief global suspension of Anthropic's Mythos-tier models under an export-control directive showed that frontier access can be cut with little notice, even for paying customers. Redundancy is now a concrete operational concern, not a theoretical one.
Cross-checking catches errors
Underwriting, claims, and compliance demand high accuracy. Running the same analysis through more than one model catches hallucinations and blind spots any single model would miss. Perplexity's cross-model mode is this idea built into a product.
Switching costs stay low
All major platforms use conversational interfaces, so the learning curve between them is small. Interoperability standards continue to lower the cost of moving work between models, which keeps lock-in risk manageable when governance is designed for it.
For insurance organizations
A practical way to match platforms to work
These mappings reflect capabilities as of the verification date and should be revisited as models and regulations evolve. They are a starting point for judgment, not a substitute for it.
| Use this | For | Because |
|---|---|---|
| Claude | Contract review, compliance writing, underwriting analysis, serious multi-step reasoning | Lowest hallucination and strongest financial reasoning, best suited to open-ended analytical work where accuracy carries regulatory weight. |
| Copilot | Grounded assistance across Microsoft 365, drafting and summarizing over your own documents | Native to the Microsoft estate with inherited governance, strong where the answer lives in your own content. |
| Perplexity | Regulatory research, competitive intelligence, market surveillance | Real-time search with inline citations and finance-focused source modes. |
| Gemini | Google Workspace-native organizations, multimodal and document analysis | Seamless where a carrier or agency already runs on Workspace, competitive on price. |
| ChatGPT | Customer-facing content, training, broad general use | Broadest ecosystem and third-party integrations. |
| Llama / DeepSeek | Data-sovereign, self-hosted, high-volume processing | Runs inside your perimeter with no data leaving it. Requires in-house capability, and DeepSeek only self-hosted under compliance review, never the hosted service. |
Capabilities are table stakes. The decision that matters is fit and governance: which tool suits the work, which plan tier the data requires, and whether the decision is even yours to make rather than the parent organization's. A leader who can reason through those three questions is doing the work this program teaches.
In creating this field guide I collaborated with Claude for research, synthesis, and drafting, and every AI-assisted claim was verified against current primary sources before publication. Model versions, pricing, and figures are perishable and were current as of the verification date shown at the top of the page. I maintain full responsibility for the content, its accuracy, and its presentation. This disclosure is made in the spirit of the transparency and disclosure practice the Accelerator teaches.