Claude 3.5 Sonnet vs GPT-4o: What Enterprises Should Compare Instead in 2026

Claude 3.5 Sonnet no longer exists as a callable model on Anthropic's API, and GPT-4o has been retired from ChatGPT entirely. Teams still searching "Claude 3.5 Sonnet vs GPT-4o" are comparing a retired model against a legacy one, so this article covers three things: where each model actually stands, how their successors (Anthropic's Claude Fable 5, Opus 5, and Sonnet 5 against OpenAI's GPT-5.6 family) compare on benchmarks, pricing, multimodality, and compliance, and how to run a selection process that survives the deprecation cycles both vendors now impose.

TLDR:

• Claude 3.5 Sonnet is retired; GPT-4o is legacy API-only, removed from ChatGPT entirely.

• Current flagships (Claude Fable 5, Sonnet 5, GPT-5.6 Sol/Terra) cluster within cents on pricing and points on benchmarks.

• Audio is the real divide: OpenAI ships native voice; Anthropic has no audio API.

• Context window size no longer differentiates; both families offer 1M+ tokens.

• Model lifespans as short as 13 months make abstraction layers and continuous evaluation non-optional.

• Internal evaluation on your own data predicts production performance; public leaderboards do not.

Where Claude 3.5 Sonnet and GPT-4o stand today

Anthropic retired both Claude 3.5 Sonnet snapshots (claude-3-5-sonnet-20240620 and claude-3-5-sonnet-20241022) from its first-party API on October 28, 2025, after notifying developers on August 13, 2025. The recommended replacement is claude-sonnet-4-6 (Anthropic model deprecations). Partner platforms followed on their own schedules, and migrations can require code changes beyond a simple model ID swap.

The standard gpt-4o remains available in the API with a 128,000-token context window, 16,384 max output tokens, and pricing of $2.50 per million input tokens and $10.00 per million output tokens. But OpenAI announced its retirement from ChatGPT on January 29, 2026 and subsequently completed the retirement across ChatGPT plans. The realtime and audio variants also have separate deprecation schedules and replacement models.

Claude 3.5 Sonnet is unavailable, while GPT-4o is now a legacy API option rather than a current flagship. The useful comparison sits between the current flagship families.

Current model lineups and pricing

Anthropic: Claude Fable 5, Opus 5, Sonnet 5, and Haiku 4.5

Anthropic describes Claude Fable 5 as "Anthropic's most capable widely released model," available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry (Anthropic models overview).

Sources: Anthropic models overview, Anthropic pricing

Two details matter for cost modeling. Claude Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 5, and Sonnet 4.6 support up to 300K output tokens on the Message Batches API via the output-300k-2026-03-24 beta header. And Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text than Sonnet 4.6 and earlier, so a per-token price comparison across generations understates the real bill.

OpenAI: the GPT-5.6 family

OpenAI announced GPT-5.6 Sol, Terra, and Luna on July 30, 2026, with official guidance to use Sol for complex reasoning and coding, Terra to balance intelligence and cost, and Luna for cost-sensitive, high-volume workloads. All three offer a 1,050,000-token context window, 128,000 max output tokens, a February 16, 2026 knowledge cutoff, and text plus image input with text output (GPT-5.6 docs).

Sol is priced at $5 input / $30 output per million tokens. The developer pricing docs list Terra at $2 / $12 and Luna at $0.20 / $1.20. Confirm the applicable price for your account and processing tier before modeling spend.

At the Sonnet 5 / Terra tier the two vendors now price within cents of each other, and both flagship families offer 1M+ token context windows. Context window size, the headline differentiator in the 3.5 Sonnet vs GPT-4o era, no longer separates them.

Benchmark comparison: close scores, incompatible methods

Read every score in this section with two caveats. First, vendor-reported figures use proprietary scaffolds while third-party figures use independent harnesses, so scores sharing a benchmark label often cannot be compared. Second, OpenAI's own audit found 59.4% of 138 audited SWE-bench Verified problems contained material issues in test design or problem descriptions, and OpenAI stopped reporting scores on it. Understanding which benchmark categories measure what matters more than any single number.

Public results still place the current Claude and OpenAI families near the frontier on graduate-level reasoning, coding, mathematics, and human preference. The exact ordering changes with the model variant, reasoning setting, scaffold, test version, and evaluation harness. Standardized coding evaluations, for example, can produce a different ordering from vendor-reported results or human-preference arenas.

The broader pattern undercuts benchmark-driven selection. Frontier scores are increasingly clustered, and older evaluations are becoming saturated or vulnerable to contamination. MMLU is effectively saturated, a problem covered in depth in Galileo's analysis of how MMLU tests the limits of language models. HumanEval, cited heavily in the original 3.5 Sonnet vs GPT-4o comparisons, no longer appears in vendor reporting for current models. 

For agentic workloads specifically, Galileo's Agent Leaderboard v2 evaluated top LLMs as real-world support agents across five industries and found no single model dominates all domains. Galileo's LLM benchmarking guide covers how to move from public scores to internal evaluation sets.

Multimodal capabilities: audio is the real divide

Vision is now table stakes. All current Claude models accept text and image input, with up to 600 images per request on 1M-context models and native PDF processing up to 32 MB and 600 pages. GPT-5.6 Sol, Terra, and Luna all accept text and image input.

Audio remains one-sided. OpenAI ships a complete native voice stack: the Realtime API processes and generates audio in a single model call over WebRTC, WebSockets, or SIP, with gpt-realtime-2.1 as the current model plus dedicated speech-to-text and text-to-speech models. Anthropic's API documentation includes no native audio input, speech output, or real-time voice; even Claude Code's voice dictation is unavailable with Anthropic API keys, Bedrock, Google Cloud, or Microsoft Foundry. Enterprise voice stacks built around Claude therefore require third-party speech-to-text and text-to-speech components. If your application talks to users, OpenAI is the shorter implementation path.

One caution applies to both vendors' 1M-token claims. Advertised window size is not effective reasoning capacity: the NoLiMa benchmark found 11 of 13 models claiming 128K+ context fell below 50% of their short-context baselines at just 32K tokens, with GPT-4o declining from 99.3% to 69.7%. Test long-context reliability on your own documents rather than trusting the spec sheet, using a structured evaluation process rather than spot checks.

Enterprise compliance, data residency, and deployment routes

Both vendors clear major enterprise certification bars, so the differences live in the fine print of each deployment route.

Claude is available via the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud's Gemini Enterprise Agent Platform, and Microsoft Foundry. Anthropic has ISO/IEC 42001 certification and offers HIPAA-ready BAAs, SSO, SCIM, audit logs via a Compliance API, and per-request US routing through the inference_geo parameter, though that parameter works only on the direct API, not Bedrock, Google Cloud, or most Foundry deployments. 

Two retention caveats matter: Claude Fable 5 is a "Covered Model" requiring 30-day retention, so zero data retention is not available for it, and trust-and-safety flags can extend retention to two years. For public-sector workloads, deployment-route approval must be checked at the cloud-platform level.

OpenAI holds SOC 2 Type 2 for the API Platform and ChatGPT Enterprise, ISO/IEC 27001, 27701, 27017, 27018, and 42001 certifications, and offers HIPAA BAAs for eligible healthcare API customers. Web Search is not HIPAA eligible. Data residency spans ten regions including the US, Europe, UK, Japan, and India, but non-US residency requires a Modified Retention amendment and residency endpoints carry a 10% pricing uplift for models released on or after March 5, 2026. Azure AI Foundry adds Data Zone deployments and additional government and regulated-workload options.

On safety governance, the philosophical split from the 3.5 Sonnet era persists in updated form. Anthropic's current constitution orders four priorities, with Claude instructed that "in cases of apparent conflict, Claude should generally prioritize these properties in the order in which they are listed," backed by a Responsible Scaling Policy and AI Safety Level protections. 

OpenAI runs a Preparedness Framework tracking biological, cyber, and self-improvement capabilities, and trains models via deliberative alignment, which directly teaches reasoning models human-written safety specifications. Teams that value documented, ordered principles may find Anthropic's written constitution easier to map into an audit narrative. Both approaches can be assessed within an enterprise AI governance program, but neither replaces workload-specific controls and evidence.

Deprecation risk and token economics

Model lifespan has become a first-order selection criterion. Anthropic's Claude Sonnet 4 and Opus 4, released May 2025, were retired by June 15, 2026, a 13-month lifespan. Anthropic commits to at least 60 days' notice; OpenAI commits to at least 6 months for GA models but as little as 2 weeks for previews, warning: "We don't recommend using preview models for business-critical production workloads unless you can migrate on short notice." API and inference-setting differences can also make migrations more involved than changing a model name.

Both vendors offer 50% batch discounts and prompt caching, with different mechanics. Anthropic charges 0.1× base input for cache hits but 1.25× for five-minute cache writes or 2× for one-hour cache writes. OpenAI's caching is automatic on supported models and does not impose the same write-fee structure on pre-GPT-5.6 models. These modifiers can dominate costs in agentic systems, where repeated context and multi-step calls multiply token usage.

How enterprises actually choose in 2026

The single-vendor decision the original Claude 3.5 Sonnet vs GPT-4o framing implied is no longer the only practical pattern. Many enterprises test multiple model families and route workloads according to modality, performance, cost, and compliance requirements rather than naming one universal winner.

A workable selection process for the current market:

1. Match modality first. Voice applications point to OpenAI's native Realtime stack. Document-heavy and coding-agent workloads should be tested directly against Claude and OpenAI models using representative tasks.

2. Map compliance by deployment route, not vendor. ZDR eligibility, data residency mechanics, and public-sector approval status differ between a vendor's direct API and its cloud-partner routes, as detailed above.

3. Model cost with caching, batch, and tokenizer effects included. List price per million tokens misleads at agent volumes.

4. Build behind an abstraction layer from day one. Short model lifespans make prompt portability and migration testing budget items, not nice-to-haves. Portable prompts and provider-neutral application interfaces reduce the cost of changing models.

5. Test on your own data. Public leaderboards cannot predict performance on proprietary enterprise knowledge or internal workflows. Enterprises increasingly rely on internal golden datasets, and for agent workloads that means agent-specific evaluation metrics like tool selection quality and action completion rather than QA benchmarks. Galileo's survey of 500+ AI practitioners found evaluation coverage gaps predict production incidents.

Model Selection Is a Continuous Process, Not a One-Time Bake-Off

The Claude 3.5 Sonnet vs GPT-4o comparison that brought most teams here is already obsolete: one model is retired, the other is legacy, and their successors cluster so tightly on benchmarks and pricing that no single number separates them. The real differentiators in 2026 are modality (OpenAI's native voice stack vs. Anthropic's document and coding strengths), compliance mechanics that vary by deployment route rather than vendor, and token economics that only become clear when caching, batch discounts, and tokenizer differences are modeled together. 

Model lifespans as short as 13 months mean today's winner can be tomorrow's migration project, which makes continuous evaluation on your own data the only durable selection strategy. Galileo provides the evaluation infrastructure that turns model comparison from a quarterly exercise into a continuous engineering process:

• Side-by-side experiments: Compare up to five model providers on the same dataset with differences in outputs, metrics, latency, and token usage surfaced automatically.

• Luna-2 evaluation models: Purpose-built SLMs that score responses at 98% lower cost than GPT-4-based evaluation with sub-200ms latency, making 100% traffic evaluation viable.

• Signals: Automatic failure detection that surfaces regression patterns, security leaks, and policy drift across production traces without manual search.

• Agentic metrics: Purpose-built measures like Tool Selection Quality and Action Completion that evaluate agent behavior where QA benchmarks cannot.

• CI/CD evaluation gates: Experiments run as unit tests in your pipeline so a model swap or version bump cannot silently regress your agents.

• Eval-to-guardrail lifecycle: Offline evals become production-enforced standards automatically, so evaluation criteria persist even as underlying models change.

Book a demo to see how Galileo turns model selection into continuous evaluation on your own production data.

Frequently Asked Questions

Is Claude 3.5 Sonnet Still Available?

No. Anthropic retired both Claude 3.5 Sonnet snapshots from its first-party API on October 28, 2025. The recommended replacement is Claude Sonnet 4.6, though Anthropic's current flagship family includes Claude Fable 5, Opus 5, and Sonnet 5. Teams still running applications against 3.5 Sonnet on partner platforms should verify availability and plan migration, as partner deprecation schedules vary.

Is GPT-4o Still Available in 2026?

The standard GPT-4o model remains available in OpenAI's API with 128K context and $2.50/$10.00 per million token pricing, but OpenAI retired it from ChatGPT entirely in early 2026. The realtime and audio variants have separate deprecation timelines. OpenAI's current flagship family is GPT-5.6, with Sol, Terra, and Luna variants covering different cost and capability tiers.

How Do Claude Fable 5 and GPT-5.6 Sol Compare?

Both sit at the frontier on graduate-level reasoning, coding, and mathematics benchmarks, but exact ordering shifts with evaluation harness, scaffold, and reasoning settings. Vendor-reported and third-party scores often use incompatible methods, making direct numerical comparison unreliable. The practical differences are in modality (OpenAI has native audio; Anthropic does not), pricing structure (Claude Fable 5 at $10/$50 vs. GPT-5.6 Sol at $5/$30), and compliance mechanics that vary by deployment route.

How Do I Choose Between Claude and GPT for Enterprise AI?

Match modality first: voice applications favor OpenAI's Realtime API, while document-heavy and coding workloads should be tested on both families using representative tasks. Map compliance requirements by deployment route rather than vendor name, since data residency, zero-data-retention eligibility, and public-sector approvals differ between direct API and cloud-partner deployments. Build behind an abstraction layer from day one, because model lifespans as short as 13 months make migration a recurring cost rather than a one-time event.

How Does Galileo Help With Model Selection and Evaluation?

Galileo turns model comparison into a continuous process rather than a one-time bake-off. Side-by-side experiments compare up to five providers on the same dataset, Luna-2 evaluation models make 100% traffic scoring affordable at sub-200ms latency, and CI/CD evaluation gates ensure a model swap cannot silently regress agent behavior. Signals surfaces failure patterns across production traces automatically, so teams catch regressions from model updates before users do.